Email alert
2023 Vol. 45, No. 1
Display Method:
2023, 45(1): .
Abstract:
2023, 45(1): 1-13.
doi: 10.11999/JEIT220183
Abstract:
With the advent of the post-Moore era, Field Programmable Gate Array (FPGA) is widely used in various fields such as Internet of Things (IoTs), 5G communication, aerospace, weapons and equipment because of its flexible repetitive programmable characteristics and low development cost. As a necessary means in the process of FPGA design and development, FPGA Electronic Design Automation (EDA) has received extensive attention from industry and academia. Especially driven by machine learning, the running efficiency and Quality of Result (QoR) have been significantly improved. A brief overview of the conceptual connotation of FPGA EDA and machine learning is presented at first, Then, the application of machine learning to different FPGA EDA stages such as High Level Synthesis (HLS), logic synthesis, placement and routing is summarized. Finally, the development of FPGA EDA technology based on machine learning is prospected. It is expected to provide reference for experts and scholars in this field and related fields, and provide technical support for the development of China’s integrated circuit industry in the post-Moore era.
With the advent of the post-Moore era, Field Programmable Gate Array (FPGA) is widely used in various fields such as Internet of Things (IoTs), 5G communication, aerospace, weapons and equipment because of its flexible repetitive programmable characteristics and low development cost. As a necessary means in the process of FPGA design and development, FPGA Electronic Design Automation (EDA) has received extensive attention from industry and academia. Especially driven by machine learning, the running efficiency and Quality of Result (QoR) have been significantly improved. A brief overview of the conceptual connotation of FPGA EDA and machine learning is presented at first, Then, the application of machine learning to different FPGA EDA stages such as High Level Synthesis (HLS), logic synthesis, placement and routing is summarized. Finally, the development of FPGA EDA technology based on machine learning is prospected. It is expected to provide reference for experts and scholars in this field and related fields, and provide technical support for the development of China’s integrated circuit industry in the post-Moore era.
2023, 45(1): 14-23.
doi: 10.11999/JEIT220391
Abstract:
Logic synthesis is a critical step in the Electronic Design Automation(EDA). Traditional global heuristic-based logic synthesis has many challenges as computing power keeps increasing and new computing paradigms emerge. There is a problem with heuristic algorithm in a suboptimal solution. As computing power improving, logic optimization is increasingly pursuing exact solutions rather than suboptimal solutions. First, the logic representations and the Boolean SATisfiability(SAT) problem are briefly described. Then, the research progress of exact synthesis in area optimization and depth optimization of Boolean logic network at two aspects, exact synthesis algorithm and encoding, are introduced. Finally, the future trends in exact synthesis are discussed.
Logic synthesis is a critical step in the Electronic Design Automation(EDA). Traditional global heuristic-based logic synthesis has many challenges as computing power keeps increasing and new computing paradigms emerge. There is a problem with heuristic algorithm in a suboptimal solution. As computing power improving, logic optimization is increasingly pursuing exact solutions rather than suboptimal solutions. First, the logic representations and the Boolean SATisfiability(SAT) problem are briefly described. Then, the research progress of exact synthesis in area optimization and depth optimization of Boolean logic network at two aspects, exact synthesis algorithm and encoding, are introduced. Finally, the future trends in exact synthesis are discussed.
2023, 45(1): 24-32.
doi: 10.11999/JEIT220350
Abstract:
Multi-core chips can provide mighty computing capability for mobile intelligent terminals, but their performance is constraint by thermal and power issues. For this problem, this paper proposes a multi-core chip dynamic power management framework based on reinforcement learning. First, based on GEM5, a dynamic voltage and frequency scaling simulation system of the multi-core chips is established. Second, a chip power model characterization method is adopted, which takes CMOS physical characteristics into consideration to realize online real-time power monitoring. Finally, a gradient reward method for the multi-core chips is designed, and a Deep Q Network (DQN) algorithm is used to learn the power management strategy for the multi-core chips. Compared with conventional Ondemand and MaxBIPS schemes, the simulation results show that the proposed framework achieves 2.12% and 4.03% improvement in computational performance of the multi-core chips respectively.
Multi-core chips can provide mighty computing capability for mobile intelligent terminals, but their performance is constraint by thermal and power issues. For this problem, this paper proposes a multi-core chip dynamic power management framework based on reinforcement learning. First, based on GEM5, a dynamic voltage and frequency scaling simulation system of the multi-core chips is established. Second, a chip power model characterization method is adopted, which takes CMOS physical characteristics into consideration to realize online real-time power monitoring. Finally, a gradient reward method for the multi-core chips is designed, and a Deep Q Network (DQN) algorithm is used to learn the power management strategy for the multi-core chips. Compared with conventional Ondemand and MaxBIPS schemes, the simulation results show that the proposed framework achieves 2.12% and 4.03% improvement in computational performance of the multi-core chips respectively.
2023, 45(1): 33-41.
doi: 10.11999/JEIT220534
Abstract:
With the increasing scarcity of medical resources and the aging of the population, cardiovascular disease has posed a great threat to human health. Portable devices with ElectroCardioGram (ECG) detection can effectively reduce the threat of cardiovascular disease to patients. In this paper, a hybrid multi-mode Convolutional Neural Network(CNN) accelerator is designed for monitoring the patient's ECG. Firstly, a one-Dimensional Convolutional Neural Network(1D-CNN) model is introduced for ECG classification, then an efficient accelerator is designed for this model, which adopts a multi-parallel expansion strategy and multi-data stream operation mode to complete the acceleration and optimization of convolution loops. The proposed operation mode can highly reuse data in time and space, and improve the utilization of hardware resources, thereby improving the hardware efficiency of the hardware accelerator. Finally, the prototype verification is completed based on the Xilinx ZC706 hardware platform. The results show 2247 LUTs and 80 DSPs are consumed. At 200 MHz operating frequency, the overall performance can reach 28.1 GOPS, and the hardware efficiency reaches 12.82 GOPS/kLUT.
With the increasing scarcity of medical resources and the aging of the population, cardiovascular disease has posed a great threat to human health. Portable devices with ElectroCardioGram (ECG) detection can effectively reduce the threat of cardiovascular disease to patients. In this paper, a hybrid multi-mode Convolutional Neural Network(CNN) accelerator is designed for monitoring the patient's ECG. Firstly, a one-Dimensional Convolutional Neural Network(1D-CNN) model is introduced for ECG classification, then an efficient accelerator is designed for this model, which adopts a multi-parallel expansion strategy and multi-data stream operation mode to complete the acceleration and optimization of convolution loops. The proposed operation mode can highly reuse data in time and space, and improve the utilization of hardware resources, thereby improving the hardware efficiency of the hardware accelerator. Finally, the prototype verification is completed based on the Xilinx ZC706 hardware platform. The results show 2247 LUTs and 80 DSPs are consumed. At 200 MHz operating frequency, the overall performance can reach 28.1 GOPS, and the hardware efficiency reaches 12.82 GOPS/kLUT.
2023, 45(1): 42-48.
doi: 10.11999/JEIT211272
Abstract:
The Physical Unclonable Function (PUF) can extract the process error of the integrated circuit during the processing and convert it into a key for security authentication. It is often used in occasions where resources and power consumption are limited, practical PUF circuits require extremely high hardware utilization efficiency and resistant to machine learning attack. A low-cost PUF circuit design scheme based on the sub-threshold current array discharge scheme is proposed. The current of the sub-threshold current array has extremely high non-linear characteristics. By introducing a gate-controlled switch and a cross-coupling structure, the uniqueness and stability of the PUF circuit can be significantly improved. In addition, the design of introducing sub-threshold current can greatly improve the security of PUF and reduce the modeling attacks of traditional attack methods. In order to improve the resource utilization of the chip, the area of the proposed PUF cell is only 377.4 μm2 through a detailed and compact layout design and optimization, making it particularly suitable for low-power and low-cost applications such as IoT. The simulation results show that the PUF of the sub-threshold circuit discharge array proposed in this paper has good uniqueness and stability, and the uniqueness is 48.85% under the standard temperature and voltage without the need to calibrate the circuit. In the temperature range of –20°C to 80°C and the voltage variation range of 0.9 V to 1.3 V, its reliability reaches 99.47%.
The Physical Unclonable Function (PUF) can extract the process error of the integrated circuit during the processing and convert it into a key for security authentication. It is often used in occasions where resources and power consumption are limited, practical PUF circuits require extremely high hardware utilization efficiency and resistant to machine learning attack. A low-cost PUF circuit design scheme based on the sub-threshold current array discharge scheme is proposed. The current of the sub-threshold current array has extremely high non-linear characteristics. By introducing a gate-controlled switch and a cross-coupling structure, the uniqueness and stability of the PUF circuit can be significantly improved. In addition, the design of introducing sub-threshold current can greatly improve the security of PUF and reduce the modeling attacks of traditional attack methods. In order to improve the resource utilization of the chip, the area of the proposed PUF cell is only 377.4 μm2 through a detailed and compact layout design and optimization, making it particularly suitable for low-power and low-cost applications such as IoT. The simulation results show that the PUF of the sub-threshold circuit discharge array proposed in this paper has good uniqueness and stability, and the uniqueness is 48.85% under the standard temperature and voltage without the need to calibrate the circuit. In the temperature range of –20°C to 80°C and the voltage variation range of 0.9 V to 1.3 V, its reliability reaches 99.47%.
2023, 45(1): 49-58.
doi: 10.11999/JEIT211462
Abstract:
The electromagnetic side channel information has the characteristics of complexity, disorder and low signal-to-noise ratio, which has a great impact on the results of side channel analysis. Based on the characteristics of electromagnetic data, in this paper, an alignment method using maximum correlation difference is proposed, which estimates the delay based on the similarity between the autocorrelation function of the reference signal and the cross-correlation function of the signal to be aligned. At the same time, a noise reduction method of K Singular Value Decomposition(KSVD) dictionary learning is proposed, which performs alternately sparse coding and dictionary update to filter out high-frequency noise. In order to verify the optimization effect of the data preprocessing method on the side channel analysis results, an electromagnetic side channel evaluation system is designed and built, and the near-field electromagnetic side channel information collection and analysis are carried out based on the actual chip. The proposed preprocessing method is used in this paper to align and reduce noise of electromagnetic data, and through t-test leakage assessment and correlation electromagnetic analysis, the maximum correlation coefficient alignment and wavelet noise reduction methods are compared, which can improve the efficiency of side-channel attacks 29.91% and 55.23%.
The electromagnetic side channel information has the characteristics of complexity, disorder and low signal-to-noise ratio, which has a great impact on the results of side channel analysis. Based on the characteristics of electromagnetic data, in this paper, an alignment method using maximum correlation difference is proposed, which estimates the delay based on the similarity between the autocorrelation function of the reference signal and the cross-correlation function of the signal to be aligned. At the same time, a noise reduction method of K Singular Value Decomposition(KSVD) dictionary learning is proposed, which performs alternately sparse coding and dictionary update to filter out high-frequency noise. In order to verify the optimization effect of the data preprocessing method on the side channel analysis results, an electromagnetic side channel evaluation system is designed and built, and the near-field electromagnetic side channel information collection and analysis are carried out based on the actual chip. The proposed preprocessing method is used in this paper to align and reduce noise of electromagnetic data, and through t-test leakage assessment and correlation electromagnetic analysis, the maximum correlation coefficient alignment and wavelet noise reduction methods are compared, which can improve the efficiency of side-channel attacks 29.91% and 55.23%.
2023, 45(1): 59-67.
doi: 10.11999/JEIT220389
Abstract:
To solve the security problems of long chip production chain, poor security and low reliability, leading to prevent Hardware Trojan (HT) detection, an HT detection method based on bypass signal analysis is proposed, by means of Linear Discriminant Analysis (LDA) classification algorithm to find the difference in time delay so as to distinguish HT. Then, the polynomial regression algorithm is used to fit the delay feature of the Trojan, and the feature library of the Trojan is established based on the regression function. The experimental results show that the proposed LDA combined with linear regression algorithm can identify HT circuits according to the delay feature, and its HT detection rate is better than other methods. Moreover, it reduces the difficulty of Trojan horse detection as the scale of the circuit increases. Through the research of this method, it has an important guiding role in identifying HT circuits and improving chip security and reliability.
To solve the security problems of long chip production chain, poor security and low reliability, leading to prevent Hardware Trojan (HT) detection, an HT detection method based on bypass signal analysis is proposed, by means of Linear Discriminant Analysis (LDA) classification algorithm to find the difference in time delay so as to distinguish HT. Then, the polynomial regression algorithm is used to fit the delay feature of the Trojan, and the feature library of the Trojan is established based on the regression function. The experimental results show that the proposed LDA combined with linear regression algorithm can identify HT circuits according to the delay feature, and its HT detection rate is better than other methods. Moreover, it reduces the difficulty of Trojan horse detection as the scale of the circuit increases. Through the research of this method, it has an important guiding role in identifying HT circuits and improving chip security and reliability.
2023, 45(1): 68-77.
doi: 10.11999/JEIT221263
Abstract:
SR Latches Physical Unclonable Functions (PUFs) are the most popular FPGA-based cryptographic applications and have a broad market in lightweight IoT devices. To realize a symmetric unbiased SR latch PUF, different implementation methods that increase area consumption have been proposed. In this paper, a novel MUX-unit-based delay gate is proposed to form the M_SR PUF unit, and the output of the SR latch in the steady state is extracted as the response of the PUF. To verify the proposed M_SR PUF, it is implemented on three series of FPGAs from Xilinx Virtex-6, Virtex-7 and Kintex-7. It is worth mentioning that the symmetrical layout is relatively simple to implement through “hard macros”, which ensures better performance of PUF. The experimental results show that the proposed M_SR PUF can work stably under an ultra-wide range of environmental changes (temperature: 0°C ~80 °C; voltage: 0.8~1.2 V) with an average uniqueness of 50.125%. Furthermore, the proposed M_SR PUF unit is characterized by low overhead, consuming only 4 MUXs and 2 DFFs, and produces a high-entropy response suitable for hardware security applications.
SR Latches Physical Unclonable Functions (PUFs) are the most popular FPGA-based cryptographic applications and have a broad market in lightweight IoT devices. To realize a symmetric unbiased SR latch PUF, different implementation methods that increase area consumption have been proposed. In this paper, a novel MUX-unit-based delay gate is proposed to form the M_SR PUF unit, and the output of the SR latch in the steady state is extracted as the response of the PUF. To verify the proposed M_SR PUF, it is implemented on three series of FPGAs from Xilinx Virtex-6, Virtex-7 and Kintex-7. It is worth mentioning that the symmetrical layout is relatively simple to implement through “hard macros”, which ensures better performance of PUF. The experimental results show that the proposed M_SR PUF can work stably under an ultra-wide range of environmental changes (temperature: 0°C ~80 °C; voltage: 0.8~1.2 V) with an average uniqueness of 50.125%. Furthermore, the proposed M_SR PUF unit is characterized by low overhead, consuming only 4 MUXs and 2 DFFs, and produces a high-entropy response suitable for hardware security applications.
2023, 45(1): 78-86.
doi: 10.11999/JEIT211579
Abstract:
For stream cipher algorithms of different granularity, reconfigurable cryptographic processors have poor compatibility and low implementation performance. In this paper, the multi-level parallelism of stream cipher algorithms is analyzed and a pre-extraction update model of the Feedback Shift Register(FSR) is established. Based on this, a Reconfigurable Feedback-shift-register Arithmetic Unit (RFAU) is proposed to apply to the cryptographic array architecture, which can be compatible with stream cipher algorithms on different Galois fields. Moreover, parallel extraction and pipeline processing strategies are executed to exploit the FSR-level parallelism of stream cipher, which effectively improve the performance of stream cryptographic algorithms on the Coarse-Grained Reconfigurable Array (CGRA) platform. The experimental results show that the performance improvement of the experimental platform brought by RFAU is reached about 23%~186% for the stream ciphers on the Galois Field (GF)(2), compared with the other reconfigurable processors. For the stream ciphers on the GF (2u) field, the throughput rate is improved to about 66%~79%, and the area efficiency is enhanced to about 64%~91%.
For stream cipher algorithms of different granularity, reconfigurable cryptographic processors have poor compatibility and low implementation performance. In this paper, the multi-level parallelism of stream cipher algorithms is analyzed and a pre-extraction update model of the Feedback Shift Register(FSR) is established. Based on this, a Reconfigurable Feedback-shift-register Arithmetic Unit (RFAU) is proposed to apply to the cryptographic array architecture, which can be compatible with stream cipher algorithms on different Galois fields. Moreover, parallel extraction and pipeline processing strategies are executed to exploit the FSR-level parallelism of stream cipher, which effectively improve the performance of stream cryptographic algorithms on the Coarse-Grained Reconfigurable Array (CGRA) platform. The experimental results show that the performance improvement of the experimental platform brought by RFAU is reached about 23%~186% for the stream ciphers on the Galois Field (GF)(2), compared with the other reconfigurable processors. For the stream ciphers on the GF (2u) field, the throughput rate is improved to about 66%~79%, and the area efficiency is enhanced to about 64%~91%.
2023, 45(1): 87-95.
doi: 10.11999/JEIT211485
Abstract:
Floating-point multipliers are the key operational units in High Dynamic Range(HDR) image processing and wireless communication systems. Compared to the fixed-point multipliers, floating-point multipliers have a higher dynamic range and also higher complexity. As a newly emerging paradigm, the hardware resources and power consumption of the circuits can be greatly reduced by approximate computing within an acceptable accuracy loss. According to the probability of 1 in the partial product array, an Approximate Floating-point Multiplier(App-Fp-Mul) is proposed in this paper. An approximate 4-2 compressor and or-gate based compression method are proposed to reduce the resource and power consumption of the floating-point multiplier with small precision loss. Compared with the accurate design, the proposed approximate floating-point multiplier can reduce the area, and power delay product by 20%, and 58% respectively when the Normalized Mean Error Distance (NMED) is 0.0014. And the proposed floating-point multiplier has higher accuracy and a smaller power delay product than the existing approximate designs with the same approximate bit width. Finally, the proposed approximate floating-point multiplier is verified with high dynamic range image processing application. The result of processing can reach 83.16 dB peak signal to noise ratio and 99.9989% structure similarity, which is obviously better than the existing approximate design.
Floating-point multipliers are the key operational units in High Dynamic Range(HDR) image processing and wireless communication systems. Compared to the fixed-point multipliers, floating-point multipliers have a higher dynamic range and also higher complexity. As a newly emerging paradigm, the hardware resources and power consumption of the circuits can be greatly reduced by approximate computing within an acceptable accuracy loss. According to the probability of 1 in the partial product array, an Approximate Floating-point Multiplier(App-Fp-Mul) is proposed in this paper. An approximate 4-2 compressor and or-gate based compression method are proposed to reduce the resource and power consumption of the floating-point multiplier with small precision loss. Compared with the accurate design, the proposed approximate floating-point multiplier can reduce the area, and power delay product by 20%, and 58% respectively when the Normalized Mean Error Distance (NMED) is 0.0014. And the proposed floating-point multiplier has higher accuracy and a smaller power delay product than the existing approximate designs with the same approximate bit width. Finally, the proposed approximate floating-point multiplier is verified with high dynamic range image processing application. The result of processing can reach 83.16 dB peak signal to noise ratio and 99.9989% structure similarity, which is obviously better than the existing approximate design.
2023, 45(1): 96-105.
doi: 10.11999/JEIT220059
Abstract:
As the Integrated Circuit (IC) industry enters the post-Moore era, the cost of one-time chip design is getting higher and higher, and threats of reverse engineering to recover the original design increase continually. In order to resist the reverse attacks, an automatic logic obfuscation technology based on genetic algorithm is proposed. By analyzing the topology network of registers in netlist, fan-in and fan-out nets are chosen to create redundant connections between registers, so that obfuscating the similarity among the word-level registers, and at last resist the reverse attacks to recover the word-level variable, control logic and dataflow. Experiments on the SM4 circuit shows, compared to the original netlist, after the proposed method’s obfuscation, the standardized mutual information of reverse attack result is reduced by 46%, and the topology complexity increased by 61.46 times. And compared to random obfuscation, the efficiency of the proposed method is increased to 2.718 times, and the area cost is reduced by 70.8%.
As the Integrated Circuit (IC) industry enters the post-Moore era, the cost of one-time chip design is getting higher and higher, and threats of reverse engineering to recover the original design increase continually. In order to resist the reverse attacks, an automatic logic obfuscation technology based on genetic algorithm is proposed. By analyzing the topology network of registers in netlist, fan-in and fan-out nets are chosen to create redundant connections between registers, so that obfuscating the similarity among the word-level registers, and at last resist the reverse attacks to recover the word-level variable, control logic and dataflow. Experiments on the SM4 circuit shows, compared to the original netlist, after the proposed method’s obfuscation, the standardized mutual information of reverse attack result is reduced by 46%, and the topology complexity increased by 61.46 times. And compared to random obfuscation, the efficiency of the proposed method is increased to 2.718 times, and the area cost is reduced by 70.8%.
2023, 45(1): 106-115.
doi: 10.11999/JEIT211435
Abstract:
Graph Convolutional Networks (GCNs) have superior performance in tasks such as social networking, ecommerce, molecular structure reasoning relative to traditional artificial intelligence algorithms, and have gained intensive attention in recent years. Unlike the independent distribution of data in Convolutional Neural Networks (CNNs), GCNs pay more attention to extract feature relationships between data, which is represented by the adjacency matrix. Therefore, the input data and operands in GCNs are much sparse and there are a large amount of data transmission, which makes it a challenge to implement an efficient GCN accelerator. Resistive Random Access Memory (ReRAM) as a new type of non-volatile memory has the advantages of high density, fast read access, near-zero leakage power and processing in-memory. Using ReRAM to accelerate CNNs has been widely studied. However, the extreme sparsity of GCNs makes it inefficiency to deploy on existing accelerators. In this work, a GCN accelerator based on ReRAM is proposed. First, the calculation and memory access characteristics of different operands in the GCN are analyzed, and a novel weight and adjacency matrix mapping policy is proposed by exploiting the intensive computing characteristic of weight and adjacency matrix, so that avoiding the excessive overhead caused by massive memory accesses; As for the extremely sparse adjacency matrix, a sub-matrix partitioning algorithm and a compression mapping scheme are proposed to minimize the GCN’s ReRAM resource requirements; Moreover, efficient processing on the sparse input feature vector with COOrdinate list (COO) compression format is provided by the proposed accelerator and the regular and efficient execution with the input feature vector are ensured. Experimental results show that the proposed work achieves 483 times speedup and 1569 times energy saving compared to CPU, and achieves 28 times speedup and consumes 168 times less energy over the GPU.
Graph Convolutional Networks (GCNs) have superior performance in tasks such as social networking, ecommerce, molecular structure reasoning relative to traditional artificial intelligence algorithms, and have gained intensive attention in recent years. Unlike the independent distribution of data in Convolutional Neural Networks (CNNs), GCNs pay more attention to extract feature relationships between data, which is represented by the adjacency matrix. Therefore, the input data and operands in GCNs are much sparse and there are a large amount of data transmission, which makes it a challenge to implement an efficient GCN accelerator. Resistive Random Access Memory (ReRAM) as a new type of non-volatile memory has the advantages of high density, fast read access, near-zero leakage power and processing in-memory. Using ReRAM to accelerate CNNs has been widely studied. However, the extreme sparsity of GCNs makes it inefficiency to deploy on existing accelerators. In this work, a GCN accelerator based on ReRAM is proposed. First, the calculation and memory access characteristics of different operands in the GCN are analyzed, and a novel weight and adjacency matrix mapping policy is proposed by exploiting the intensive computing characteristic of weight and adjacency matrix, so that avoiding the excessive overhead caused by massive memory accesses; As for the extremely sparse adjacency matrix, a sub-matrix partitioning algorithm and a compression mapping scheme are proposed to minimize the GCN’s ReRAM resource requirements; Moreover, efficient processing on the sparse input feature vector with COOrdinate list (COO) compression format is provided by the proposed accelerator and the regular and efficient execution with the input feature vector are ensured. Experimental results show that the proposed work achieves 483 times speedup and 1569 times energy saving compared to CPU, and achieves 28 times speedup and consumes 168 times less energy over the GPU.
2023, 45(1): 116-124.
doi: 10.11999/JEIT211249
Abstract:
In this paper, an Integrate-and-Fire (IF) analog readout neuron circuit is proposed for Spiking Convolutional Neural Network (SCNN) based on flash array. The circuit realizes the following functions: bit line voltage clamping, current readout, current subtraction, and integrate-and-fire. A current readout method is proposed to improve the current readout range and speed by increasing by-pass current. To avoid the loss of array information caused by the traditional analog neuron reset scheme, a reset scheme with subtracting threshold voltage is proposed, which improves the integrity of information and the accuracy of the neural network. The circuit is implemented in 55 nm Complementary Metal Oxide Semiconductor (CMOS) process. Simulation results show that when output current is 20 μA and 0 μA, the read speed can be accelerated 100% and 263.6% respectively; The neuron circuit works well. And test results show that, in the current output range of 0~20 μA, the clamp voltage error is less than 0.2 mV and the fluctuation is less than 0.4 mV; The linearity of current subtraction can reach 99.9%. To study the performance of the analog neuron circuit, LeNet and AlexNet algorithm with circuit model for the recognition of the MNIST and CIFAR-10 database is tested. Test results illustrate that the neural network accuracy is improved by 1.4% and 38.8%.
In this paper, an Integrate-and-Fire (IF) analog readout neuron circuit is proposed for Spiking Convolutional Neural Network (SCNN) based on flash array. The circuit realizes the following functions: bit line voltage clamping, current readout, current subtraction, and integrate-and-fire. A current readout method is proposed to improve the current readout range and speed by increasing by-pass current. To avoid the loss of array information caused by the traditional analog neuron reset scheme, a reset scheme with subtracting threshold voltage is proposed, which improves the integrity of information and the accuracy of the neural network. The circuit is implemented in 55 nm Complementary Metal Oxide Semiconductor (CMOS) process. Simulation results show that when output current is 20 μA and 0 μA, the read speed can be accelerated 100% and 263.6% respectively; The neuron circuit works well. And test results show that, in the current output range of 0~20 μA, the clamp voltage error is less than 0.2 mV and the fluctuation is less than 0.4 mV; The linearity of current subtraction can reach 99.9%. To study the performance of the analog neuron circuit, LeNet and AlexNet algorithm with circuit model for the recognition of the MNIST and CIFAR-10 database is tested. Test results illustrate that the neural network accuracy is improved by 1.4% and 38.8%.
2023, 45(1): 125-133.
doi: 10.11999/JEIT211150
Abstract:
The improvement of application efficiency of heterogeneous computing systems is highly dependent on effective scheduling algorithms. A new list scheduling algorithm called Improved Predict Priority and Optimistic processor Selection Scheduling (IPPOSS) is proposed by this paper. By introducing the backward prediction cost of tasks in task prioritizing phase, the scheduling length is reduced. Compared with the existing work, an Improved Predict Cost Matrix (IPCM) is adopted to prioritize tasks more reasonably and a better solution in processor selection phase when keeping quadratic time complexity is obtain. IPCM, which considers various calculation and communication factors in the task prioritization stage, is easier to obtain a reasonable priority list than Predict Cost Matrix (PCM) proposed by Predict Priority Task Scheduling (PPTS). That the performance of IPPOSS is better than related algorithms is shown by the analysis of the experimental results of randomly generated application Directed Acyclic Graphs (DAGs) and real-world application DAGs.
The improvement of application efficiency of heterogeneous computing systems is highly dependent on effective scheduling algorithms. A new list scheduling algorithm called Improved Predict Priority and Optimistic processor Selection Scheduling (IPPOSS) is proposed by this paper. By introducing the backward prediction cost of tasks in task prioritizing phase, the scheduling length is reduced. Compared with the existing work, an Improved Predict Cost Matrix (IPCM) is adopted to prioritize tasks more reasonably and a better solution in processor selection phase when keeping quadratic time complexity is obtain. IPCM, which considers various calculation and communication factors in the task prioritization stage, is easier to obtain a reasonable priority list than Predict Cost Matrix (PCM) proposed by Predict Priority Task Scheduling (PPTS). That the performance of IPPOSS is better than related algorithms is shown by the analysis of the experimental results of randomly generated application Directed Acyclic Graphs (DAGs) and real-world application DAGs.
2023, 45(1): 134-149.
doi: 10.11999/JEIT211164
Abstract:
With the progress of modern communication technology, especially the rapid development of 4G, 5G and other wireless mobile communications, modulation methods with high spectrum efficiency such as multi-Quadrature Amplitude Modulation (QAM) have been widely used, which puts forward higher and stricter linear requirements for wireless communication systems. The Radio Frequency Low Noise Amplifier(RF LNA)is the first active device of the RF Front-End Module(RF FEM), and the signal quality and dynamic range of the system are directly affected by the the nonlinear characteristics of the LNA. Taking the 3rd-order intermodulation as an example, a sufficient input 3rd-order intercept point is required in LNA to ensure expected performance even with strong interfering signals. Based on the 3rd-order nonlinear model, in this article, the theoretical model of the 3rd-order intermodulation is analyzed briefly, the methods to improve the 3rd-order intercept point are sorted out, the relevant research results and progress in recent years are summarized and studied , and the future development trend is prospected.
With the progress of modern communication technology, especially the rapid development of 4G, 5G and other wireless mobile communications, modulation methods with high spectrum efficiency such as multi-Quadrature Amplitude Modulation (QAM) have been widely used, which puts forward higher and stricter linear requirements for wireless communication systems. The Radio Frequency Low Noise Amplifier(RF LNA)is the first active device of the RF Front-End Module(RF FEM), and the signal quality and dynamic range of the system are directly affected by the the nonlinear characteristics of the LNA. Taking the 3rd-order intermodulation as an example, a sufficient input 3rd-order intercept point is required in LNA to ensure expected performance even with strong interfering signals. Based on the 3rd-order nonlinear model, in this article, the theoretical model of the 3rd-order intermodulation is analyzed briefly, the methods to improve the 3rd-order intercept point are sorted out, the relevant research results and progress in recent years are summarized and studied , and the future development trend is prospected.
2023, 45(1): 150-157.
doi: 10.11999/JEIT211144
Abstract:
A Micro-Electro-Mechanical System (MEMS) electric field sensor with low driving voltage based on Lead Zirconate Titanate (PZT) is proposed. Based on the charge-induction principle, the sensitive units are composed of fixed electrodes and movable electrodes. All the fixed and movable electrodes work as sensing electrodes, in the meantime, they are also mutually shielding electrodes. Driven by the piezoelectric material PZT, the movable electrodes vibrate perpendicularly to the substrate of the sensitive chip, and they are mutually shielded with the fixed electrodes. When there is an electric field to be measured, induced current signals with a phase difference of 180° are generated respectively on the movable electrodes and the fixed electrodes. The design and finite element simulation of the sensor are carried out, the fabrication process of the sensitive microstructure is proposed, the MEMS process compatible preparation technology of movable electrode based on PZT piezoelectric material is broken through, the microsensor chip is successfully produced, and the performance of the sensor is tested. The sensor has the advantage of low working voltage. Experimental results show that, with 1 V AC driving voltage, the sensitivity of the electric field sensor system is 0.292 mV/(kV/m) and the linearity is 2.89% in the range of 0~50 kV/m electric field intensity.
A Micro-Electro-Mechanical System (MEMS) electric field sensor with low driving voltage based on Lead Zirconate Titanate (PZT) is proposed. Based on the charge-induction principle, the sensitive units are composed of fixed electrodes and movable electrodes. All the fixed and movable electrodes work as sensing electrodes, in the meantime, they are also mutually shielding electrodes. Driven by the piezoelectric material PZT, the movable electrodes vibrate perpendicularly to the substrate of the sensitive chip, and they are mutually shielded with the fixed electrodes. When there is an electric field to be measured, induced current signals with a phase difference of 180° are generated respectively on the movable electrodes and the fixed electrodes. The design and finite element simulation of the sensor are carried out, the fabrication process of the sensitive microstructure is proposed, the MEMS process compatible preparation technology of movable electrode based on PZT piezoelectric material is broken through, the microsensor chip is successfully produced, and the performance of the sensor is tested. The sensor has the advantage of low working voltage. Experimental results show that, with 1 V AC driving voltage, the sensitivity of the electric field sensor system is 0.292 mV/(kV/m) and the linearity is 2.89% in the range of 0~50 kV/m electric field intensity.
2023, 45(1): 158-167.
doi: 10.11999/JEIT211281
Abstract:
In view of the problem that the change of the density of the label distribution will lead to the change of the impedance matching relationship between the antenna and the load, and then affect the system performance. Based on the theory of electromagnetic wave propagation and the working principle of Radio Frequency IDentification (RFID), the communication link models of RFID system with sparse and dense tags are derived. Based on the transformer model and the two-port network analysis method, the mutual impedance expression of the tag antenna in the dense distribution state is derived. Using power transmission coefficient and backscattering modulation factor, the influence of tag density on RFID system performance is analyzed. Based on the principle of loading bar matching, an optimal design method for label antenna is proposed. Simulation and actual measurement results show that the performance of the improved tag is 16% higher than that of the prototype tag when the tags are densely distributed. When the tags are sparsely distributed, the performance of the improved tags reaches 96% of that of the prototype tags.
In view of the problem that the change of the density of the label distribution will lead to the change of the impedance matching relationship between the antenna and the load, and then affect the system performance. Based on the theory of electromagnetic wave propagation and the working principle of Radio Frequency IDentification (RFID), the communication link models of RFID system with sparse and dense tags are derived. Based on the transformer model and the two-port network analysis method, the mutual impedance expression of the tag antenna in the dense distribution state is derived. Using power transmission coefficient and backscattering modulation factor, the influence of tag density on RFID system performance is analyzed. Based on the principle of loading bar matching, an optimal design method for label antenna is proposed. Simulation and actual measurement results show that the performance of the improved tag is 16% higher than that of the prototype tag when the tags are densely distributed. When the tags are sparsely distributed, the performance of the improved tags reaches 96% of that of the prototype tags.
2023, 45(1): 168-180.
doi: 10.11999/JEIT211291
Abstract:
Gap Waveguide (GW) is a new kind of artificial ElectroMagnetic (EM) material based on contactless Electromagnetic Band Gap(EBG) structure. The unique contactless structure and wide EM forbidden band of GW show great advantages and flexibility in developing new EM transmission lines and shielding structures, providing a new research perspective and realization approach for microwave & millimeter-wave components, circuits and antennas, etc., and causing much attentions in recent years. Firstly, the concept and principle of GW are briefly introduced, and its technical advantages are analyzed. Then, the research progress of GW is comprehensively summarized, according to the classification of various research and application fields. In the end, the application prospect of GW in space microwave & millimeter-wave technology is discussed combining with space technology background and development needs, and the contactless suppression method of passive intermodulation and stacked-integrated technology of millimeter-wave systems are proposed. This work can provide some valuable references for the research and application of GW.
Gap Waveguide (GW) is a new kind of artificial ElectroMagnetic (EM) material based on contactless Electromagnetic Band Gap(EBG) structure. The unique contactless structure and wide EM forbidden band of GW show great advantages and flexibility in developing new EM transmission lines and shielding structures, providing a new research perspective and realization approach for microwave & millimeter-wave components, circuits and antennas, etc., and causing much attentions in recent years. Firstly, the concept and principle of GW are briefly introduced, and its technical advantages are analyzed. Then, the research progress of GW is comprehensively summarized, according to the classification of various research and application fields. In the end, the application prospect of GW in space microwave & millimeter-wave technology is discussed combining with space technology background and development needs, and the contactless suppression method of passive intermodulation and stacked-integrated technology of millimeter-wave systems are proposed. This work can provide some valuable references for the research and application of GW.
2023, 45(1): 181-190.
doi: 10.11999/JEIT211452
Abstract:
The antenna array composed of UAVs (Unmanned Aerial Vehicle Swarm) often presents near-field and sparse characteristics, which can not be adapted by the classical beamforming theory. Therefore, the near-field uniform linear array signal model is first constructed, and a simplified implementation method of near-field beamforming based on linear frequency modulation pulse compression by Taylor expansion of the signal phase difference function is proposed. Then the characteristics of near-field beamforming are analyzed in the two-dimensional space-frequency domain. From the perspective of spatial under sampling, the analytical expression of near-field beam grating lobe distribution under the condition of sparse array elements is proposed. From the perspective of spatial frequency offset and bandwidth mismatch, the variation characteristics of azimuth gain and range gain of near-field beam are analyzed and obtained. Simulation results show the effectiveness of the conclusions, and provide a theoretical support for using beamforming technology to improve the communication, reconnaissance and jamming capabilities of UAVs.
The antenna array composed of UAVs (Unmanned Aerial Vehicle Swarm) often presents near-field and sparse characteristics, which can not be adapted by the classical beamforming theory. Therefore, the near-field uniform linear array signal model is first constructed, and a simplified implementation method of near-field beamforming based on linear frequency modulation pulse compression by Taylor expansion of the signal phase difference function is proposed. Then the characteristics of near-field beamforming are analyzed in the two-dimensional space-frequency domain. From the perspective of spatial under sampling, the analytical expression of near-field beam grating lobe distribution under the condition of sparse array elements is proposed. From the perspective of spatial frequency offset and bandwidth mismatch, the variation characteristics of azimuth gain and range gain of near-field beam are analyzed and obtained. Simulation results show the effectiveness of the conclusions, and provide a theoretical support for using beamforming technology to improve the communication, reconnaissance and jamming capabilities of UAVs.
2023, 45(1): 191-199.
doi: 10.11999/JEIT211244
Abstract:
Focusing on the optimal tracking problem of multiple extended targets, a sensor control method is proposed, which can comprehensively optimize the tracking performance of multiple extended targets in the framework of FInite Set STatistics (FISST). First, a Weighted Generalized Optimal Sub-Pattern Assignment (WGOSPA) distance is proposed to construct multiple extended targets tracking multiple feature estimation of generalized dispersion around its statistical average. In addition, an optimal decision-making method of sensor control through the multi-characteristic fusion is studied and proposed. Furthermore, the numerical solution method of the optimal decision-making process of sensor control is studied by using Sequential Monte Carlo (SMC) technology. Then, the proposed sensor control strategy is realized by using Gamma Gaussian Inverse Wishart Multi-Bernoulli (GGIW-MBer) filter. Finally, the effectiveness of the proposed algorithm is verified by simulation experiments.
Focusing on the optimal tracking problem of multiple extended targets, a sensor control method is proposed, which can comprehensively optimize the tracking performance of multiple extended targets in the framework of FInite Set STatistics (FISST). First, a Weighted Generalized Optimal Sub-Pattern Assignment (WGOSPA) distance is proposed to construct multiple extended targets tracking multiple feature estimation of generalized dispersion around its statistical average. In addition, an optimal decision-making method of sensor control through the multi-characteristic fusion is studied and proposed. Furthermore, the numerical solution method of the optimal decision-making process of sensor control is studied by using Sequential Monte Carlo (SMC) technology. Then, the proposed sensor control strategy is realized by using Gamma Gaussian Inverse Wishart Multi-Bernoulli (GGIW-MBer) filter. Finally, the effectiveness of the proposed algorithm is verified by simulation experiments.
2023, 45(1): 200-207.
doi: 10.11999/JEIT211290
Abstract:
To overcome the beam split effect in ultra-bandwidth terahertz communications, the True Time Delay (TTD)-based large array antenna structures have been designed, while the power consumption and the hardware complexity both are high. To solve this problem, a serial equally spaced TTD-based sparse radio frequency chain antenna structure is proposed. Based on the designed antenna structure, the beams at all subcarriers can be changed by jointly optimizing the time delays of TTD devices and the phase of phase shifters, obtaining the beam spreading and beam convergence for serving different distributed users. Specifically, to serve users located in different directions, the frequency-dependent beams can be spread to different directions by optimizing time delays and phase shifts to realize beam spreading. Similarly, to serve users located in a single direction, the beams at all subcarriers can be changed to align with the same direction to obtain beam convergence. Finally, simulation results demonstrate that effectiveness of the designed structure and the proposed optimization scheme.
To overcome the beam split effect in ultra-bandwidth terahertz communications, the True Time Delay (TTD)-based large array antenna structures have been designed, while the power consumption and the hardware complexity both are high. To solve this problem, a serial equally spaced TTD-based sparse radio frequency chain antenna structure is proposed. Based on the designed antenna structure, the beams at all subcarriers can be changed by jointly optimizing the time delays of TTD devices and the phase of phase shifters, obtaining the beam spreading and beam convergence for serving different distributed users. Specifically, to serve users located in different directions, the frequency-dependent beams can be spread to different directions by optimizing time delays and phase shifts to realize beam spreading. Similarly, to serve users located in a single direction, the beams at all subcarriers can be changed to align with the same direction to obtain beam convergence. Finally, simulation results demonstrate that effectiveness of the designed structure and the proposed optimization scheme.
2023, 45(1): 208-217.
doi: 10.11999/JEIT211276
Abstract:
In recent years, deep learning has become one of the key technologies in the field of wireless communication. In a series of MIMO signal detection algorithms based on deep learning, most of them do not fully consider the interference cancellation problem between adjacent antennas, hence the impact of multi-user interference on the bit error rate performance can not be completely eliminated. To this end, a method that combines deep learning and Successive Interference Cancellation (SIC) algorithms for uplink signal detection in a massive MIMO system is propesed. Firstly, by optimizing the traditional Detection Network (DetNet) and improving the ScNet (Sparsely connected neural Network), a detection algorithm based on the Deep Neural Network (DNN), called Improved ScNet (ImpScNet), is proposed. On this basis, the SIC is applied to the design of the deep learning framework structure, and a massive MIMO multi-user SIC detection algorithm based on deep learning is proposed, which is called ImpScNet-SIC. This algorithm is divided into two stages on each detection layer. The first stage is provided by the ImpScNet algorithm proposed in this paper to provide the initial solution, and then the initial solution is demodulated to the corresponding constellation point as the input of the SIC, which constitutes the second stage. In addition, the ImpScNet algorithm is also used in SIC to estimate the transmitted symbols in order to obtain the best performance. Simulation results show that, compared with various typical representative algorithms, the ImpScNet-SIC detection algorithm proposed in this paper is particularly suitable for the massive MIMO signal detection. It has the advantages of fast convergence speed, stable convergence and relatively low complexity. And there is at least 0.5 dB gain in 10–3 bit error rate.
In recent years, deep learning has become one of the key technologies in the field of wireless communication. In a series of MIMO signal detection algorithms based on deep learning, most of them do not fully consider the interference cancellation problem between adjacent antennas, hence the impact of multi-user interference on the bit error rate performance can not be completely eliminated. To this end, a method that combines deep learning and Successive Interference Cancellation (SIC) algorithms for uplink signal detection in a massive MIMO system is propesed. Firstly, by optimizing the traditional Detection Network (DetNet) and improving the ScNet (Sparsely connected neural Network), a detection algorithm based on the Deep Neural Network (DNN), called Improved ScNet (ImpScNet), is proposed. On this basis, the SIC is applied to the design of the deep learning framework structure, and a massive MIMO multi-user SIC detection algorithm based on deep learning is proposed, which is called ImpScNet-SIC. This algorithm is divided into two stages on each detection layer. The first stage is provided by the ImpScNet algorithm proposed in this paper to provide the initial solution, and then the initial solution is demodulated to the corresponding constellation point as the input of the SIC, which constitutes the second stage. In addition, the ImpScNet algorithm is also used in SIC to estimate the transmitted symbols in order to obtain the best performance. Simulation results show that, compared with various typical representative algorithms, the ImpScNet-SIC detection algorithm proposed in this paper is particularly suitable for the massive MIMO signal detection. It has the advantages of fast convergence speed, stable convergence and relatively low complexity. And there is at least 0.5 dB gain in 10–3 bit error rate.
2023, 45(1): 218-226.
doi: 10.11999/JEIT211287
Abstract:
With the rise of Internet social platforms and the popularization of mobile smart terminal devices, people's demand for high-quality and real-time data has risen sharply, especially for video services such as short videos and live streams. At the same time, too many terminal devices connected to the core network increase the load of the backhaul link, so that the traditional cloud computing is difficult to meet the low-latency requirements of users for video services. By deploying edge nodes with computing and storage capabilities at the edge of the network, Mobile Edge Computing (MEC) can calculate and store closer to the users, which will reduce the data transmission delay and alleviate the network congestion. Therefore, making full use of the computing and storage resources at the edge of the network under MEC, a video request prediction method and a cooperative caching strategy based on federated learning are proposed. By federally training the proposed model Deep Request Prediction Network (DRPN) with multiple edge nodes, the video requests in the future can be predicted and then cache decisions can be made cooperatively. The simulation results show that compared with other strategies, the proposed strategy can not only effectively improve the cache hit rate and reduce the user waiting delay, but also reduce the communication cost and cache cost of the whole system to a certain extent.
With the rise of Internet social platforms and the popularization of mobile smart terminal devices, people's demand for high-quality and real-time data has risen sharply, especially for video services such as short videos and live streams. At the same time, too many terminal devices connected to the core network increase the load of the backhaul link, so that the traditional cloud computing is difficult to meet the low-latency requirements of users for video services. By deploying edge nodes with computing and storage capabilities at the edge of the network, Mobile Edge Computing (MEC) can calculate and store closer to the users, which will reduce the data transmission delay and alleviate the network congestion. Therefore, making full use of the computing and storage resources at the edge of the network under MEC, a video request prediction method and a cooperative caching strategy based on federated learning are proposed. By federally training the proposed model Deep Request Prediction Network (DRPN) with multiple edge nodes, the video requests in the future can be predicted and then cache decisions can be made cooperatively. The simulation results show that compared with other strategies, the proposed strategy can not only effectively improve the cache hit rate and reduce the user waiting delay, but also reduce the communication cost and cache cost of the whole system to a certain extent.
2023, 45(1): 227-234.
doi: 10.11999/JEIT211262
Abstract:
Considering the non-negligible communication cost problem caused by redundant gradient interactive communication between a large number of device nodes in the Federated Learning(FL) process in the Internet of Things (IoTs) scenario, gradient communication compression mechanism with adaptive threshold is proposed. Firstly, a structure of Communication-Efficient EDge-Federated Learning (CE-EDFL) is used to prevent device-side data privacy leakage. The edge server acts as an intermediary device to perform device-side local model aggregation, and the cloud performs edge server model aggregation and new parameter delivery. Secondly, in order to reduce further the communication overhead during federated learning detection, a threshold Adaptive Lazily Aggregated Gradient (ALAG) is proposed, which reduces the redundant communication between the device end and the edge server by compressing the gradient parameters of the local model. The experimental results show that the proposed algorithm can effectively improve the overall communication efficiency of the model by reducing the number of gradient interactions while ensuring the accuracy of deep learning tasks in the large-scale IoT device scenario.
Considering the non-negligible communication cost problem caused by redundant gradient interactive communication between a large number of device nodes in the Federated Learning(FL) process in the Internet of Things (IoTs) scenario, gradient communication compression mechanism with adaptive threshold is proposed. Firstly, a structure of Communication-Efficient EDge-Federated Learning (CE-EDFL) is used to prevent device-side data privacy leakage. The edge server acts as an intermediary device to perform device-side local model aggregation, and the cloud performs edge server model aggregation and new parameter delivery. Secondly, in order to reduce further the communication overhead during federated learning detection, a threshold Adaptive Lazily Aggregated Gradient (ALAG) is proposed, which reduces the redundant communication between the device end and the edge server by compressing the gradient parameters of the local model. The experimental results show that the proposed algorithm can effectively improve the overall communication efficiency of the model by reducing the number of gradient interactions while ensuring the accuracy of deep learning tasks in the large-scale IoT device scenario.
2023, 45(1): 235-242.
doi: 10.11999/JEIT211160
Abstract:
In a cellular network scenario that supports direct Vehicle-to-Vehicle(V2V) communication, the resource allocation problem of the uplink of multiple Vehicle-to-Infrastructure (V2I) in a dense environment is used. Under the interference of V2V, the slow fading statistics of the Channel State Information (CSI) of the mobile link is used, joint communication reliability, power control, an optimization model is established that maximizes the V2I channel capacity to meet the needs of heterogeneous vehicle network services. Based on this, a resource allocation algorithm based on hypergraph theory and genetic algorithm is proposed. The simulation results show that the algorithm improves the channel capacity of V2I under the premise of ensuring the reliability of V2V communication.
In a cellular network scenario that supports direct Vehicle-to-Vehicle(V2V) communication, the resource allocation problem of the uplink of multiple Vehicle-to-Infrastructure (V2I) in a dense environment is used. Under the interference of V2V, the slow fading statistics of the Channel State Information (CSI) of the mobile link is used, joint communication reliability, power control, an optimization model is established that maximizes the V2I channel capacity to meet the needs of heterogeneous vehicle network services. Based on this, a resource allocation algorithm based on hypergraph theory and genetic algorithm is proposed. The simulation results show that the algorithm improves the channel capacity of V2I under the premise of ensuring the reliability of V2V communication.
2023, 45(1): 243-253.
doi: 10.11999/JEIT211123
Abstract:
To solve the problems of signal distortion at the transceivers, information leakage and transmission interruption of users in multi-cell multi-user heterogeneous networks, a robust secure resource allocation algorithm is proposed with distortion noises. Considering the constraints of the minimum secure rate of each small-cell user, the maximum transmit power of each base station and the cross-tier interference power of each macro-cell user, an energy efficiency maximization-based resource allocation model is formulated under bounded channel uncertainties. The original non-convex optimization problem is converted into an equivalent convex one by using the Dinkelbach’s method, the worst-case method and the successive convex approximation method. And the analytical solution is obtained by using Lagrangian dual function method. Simulation results demonstrate that the proposed algorithm has better energy efficiency and robustness by comparing it with the existing algorithms.
To solve the problems of signal distortion at the transceivers, information leakage and transmission interruption of users in multi-cell multi-user heterogeneous networks, a robust secure resource allocation algorithm is proposed with distortion noises. Considering the constraints of the minimum secure rate of each small-cell user, the maximum transmit power of each base station and the cross-tier interference power of each macro-cell user, an energy efficiency maximization-based resource allocation model is formulated under bounded channel uncertainties. The original non-convex optimization problem is converted into an equivalent convex one by using the Dinkelbach’s method, the worst-case method and the successive convex approximation method. And the analytical solution is obtained by using Lagrangian dual function method. Simulation results demonstrate that the proposed algorithm has better energy efficiency and robustness by comparing it with the existing algorithms.
2023, 45(1): 254-261.
doi: 10.11999/JEIT211165
Abstract:
Driven by the rapid development of the Internet of Things, mobile Ad Hoc cloud computing and Energy Harvesting (EH) technologies meet the needs of data processing by sharing the idle resources of adjacent devices, and realize green communication. However, due to the time-varying nature of Ad Hoc cloud networks and the stochastic instability of EH, the research on reasonable task offloading schemes faces severe challenges. Considering the above problems, a distributed dynamic offloading scheme is proposed in this paper by using Lyapunov optimization theory and game theory. It is impossible for rational terminal equipment to serve other terminal equipment for free, in order to encourage terminal equipment to participate actively in the calculation offloading process, an incentive mechanism based on dynamic quotation is proposed in this paper. Compared with the existing schemes, the simulation results show that the proposed scheme can effectively improve system revenue, stabilize battery energy and reduce task queue backlog.
Driven by the rapid development of the Internet of Things, mobile Ad Hoc cloud computing and Energy Harvesting (EH) technologies meet the needs of data processing by sharing the idle resources of adjacent devices, and realize green communication. However, due to the time-varying nature of Ad Hoc cloud networks and the stochastic instability of EH, the research on reasonable task offloading schemes faces severe challenges. Considering the above problems, a distributed dynamic offloading scheme is proposed in this paper by using Lyapunov optimization theory and game theory. It is impossible for rational terminal equipment to serve other terminal equipment for free, in order to encourage terminal equipment to participate actively in the calculation offloading process, an incentive mechanism based on dynamic quotation is proposed in this paper. Compared with the existing schemes, the simulation results show that the proposed scheme can effectively improve system revenue, stabilize battery energy and reduce task queue backlog.
2023, 45(1): 262-271.
doi: 10.11999/JEIT211261
Abstract:
For the problem of Service Function Chain (SFC) anomalies due to hardware and software anomalies in network slicing scenarios, a Distributed Generative Adversarial Network (GAN)-based Time Series anomaly detection model (DTSGAN) is proposed. First, to learn the characteristics of normal data in SFC, a distributed GAN architecture is proposed for anomaly detection of multiple Virtual Network Functions (VNFs) contained in SFC. Then, a feature extractor based on sliding window data is constructed for time series data, and the feature sequence is obtained by extracting two derived characteristics and eight statistical features of the data to mine the deep-level features. Finally, in order to learn and reconstruct data characteristics, a three-layer codec constructed by Time Convolutional Network (TCN) and Auto-Encoder (AE) is proposed as a distributed generator, which measures the difference between reconstructed data and input data by anomaly score function to detect the state of VNF, and then completes the anomaly detection of SFC. The effectiveness and stability of the proposed model are verified on the dataset Clearwater using four evaluation metrics: accuracy, precision, recall and F1 score.
For the problem of Service Function Chain (SFC) anomalies due to hardware and software anomalies in network slicing scenarios, a Distributed Generative Adversarial Network (GAN)-based Time Series anomaly detection model (DTSGAN) is proposed. First, to learn the characteristics of normal data in SFC, a distributed GAN architecture is proposed for anomaly detection of multiple Virtual Network Functions (VNFs) contained in SFC. Then, a feature extractor based on sliding window data is constructed for time series data, and the feature sequence is obtained by extracting two derived characteristics and eight statistical features of the data to mine the deep-level features. Finally, in order to learn and reconstruct data characteristics, a three-layer codec constructed by Time Convolutional Network (TCN) and Auto-Encoder (AE) is proposed as a distributed generator, which measures the difference between reconstructed data and input data by anomaly score function to detect the state of VNF, and then completes the anomaly detection of SFC. The effectiveness and stability of the proposed model are verified on the dataset Clearwater using four evaluation metrics: accuracy, precision, recall and F1 score.
2023, 45(1): 272-281.
doi: 10.11999/JEIT211294
Abstract:
To address the problem that in-vehicle cyber communication messages are easily captured, an Endogenous Security Mechanism for in-Vehicle Networks (ESM-VN) based on Dynamic Heterogeneous Redundancy (DHR) architecture is proposed. Firstly, the model of vehicle in network replay attack is analyzed, the network characteristics of replay attack are summarized. Then, the implementation mechanism of DHR of vehicle network communication message is designed by using the theory of network space endogenous security to realize the coordination and unification of attack perception and active defense through dynamic adjudication and negative feedback mechanism. Analysis and simulation results show that compared with the traditional in-vehicle network defense method, the proposed mechanism can reduce the response delay by at least 50% and improve effectively the defense capability of in-vehicle network against replay attack.
To address the problem that in-vehicle cyber communication messages are easily captured, an Endogenous Security Mechanism for in-Vehicle Networks (ESM-VN) based on Dynamic Heterogeneous Redundancy (DHR) architecture is proposed. Firstly, the model of vehicle in network replay attack is analyzed, the network characteristics of replay attack are summarized. Then, the implementation mechanism of DHR of vehicle network communication message is designed by using the theory of network space endogenous security to realize the coordination and unification of attack perception and active defense through dynamic adjudication and negative feedback mechanism. Analysis and simulation results show that compared with the traditional in-vehicle network defense method, the proposed mechanism can reduce the response delay by at least 50% and improve effectively the defense capability of in-vehicle network against replay attack.
2023, 45(1): 282-290.
doi: 10.11999/JEIT211107
Abstract:
The GEostationary Orbit(GEO) GF-4 satellite has the ability to observe continuously moving ships at sea. Ship targets are often weak in the optical remote sensing images of GF-4 satellite, making it difficult to detect directly. By analyzing the wake characteristics of moving ships, a moving ship detection method based on Multi-scale Dual-neighborhood Saliency Model (MDSM) is proposed. First, the saliency of the image is calculated based on MDSM. Then, the position of the moving ship is extracted by adaptive segmentation threshold. Finally, the shape of the candidate target is verified to remove further the false target. Experimental results and analysis show that the proposed method can effectively detect multiple moving targets in GF-4 satellite images, and has better detection performance compared with the current mainstream visual saliency algorithms.
The GEostationary Orbit(GEO) GF-4 satellite has the ability to observe continuously moving ships at sea. Ship targets are often weak in the optical remote sensing images of GF-4 satellite, making it difficult to detect directly. By analyzing the wake characteristics of moving ships, a moving ship detection method based on Multi-scale Dual-neighborhood Saliency Model (MDSM) is proposed. First, the saliency of the image is calculated based on MDSM. Then, the position of the moving ship is extracted by adaptive segmentation threshold. Finally, the shape of the candidate target is verified to remove further the false target. Experimental results and analysis show that the proposed method can effectively detect multiple moving targets in GF-4 satellite images, and has better detection performance compared with the current mainstream visual saliency algorithms.
2023, 45(1): 291-299.
doi: 10.11999/JEIT211188
Abstract:
Multi-exposure image fusion aims to fuse a series of images with different exposures for the same scene, and it is the main-stream method for high dynamic range imaging. To obtain more realistic results, a Multi-Exposure image Fusion Network(MEF-Net) based on deep guided and self-learning is proposed. This network is designed to fuse any number of images with different exposures in an end-to-end way, and generate the best-fused results in an unsupervised way. In terms of the loss function, an intensity fidelity constraint term and the weighted Multi-Exposure image Fusion Structural SIMilarity(MEF-SSIM) are introduced to improve the fusion quality. Moreover, a self-learning method is adopted to fine-tune and optimize the pre-learned model, considering the fusion problem of two images under extreme exposure to mitigate the halo phenomenon generated by fusion. Based on abundant testing data, experimental results show that the proposed algorithm outperforms other mainstream methods in terms of both quantitative measurement and visual fused quality.
Multi-exposure image fusion aims to fuse a series of images with different exposures for the same scene, and it is the main-stream method for high dynamic range imaging. To obtain more realistic results, a Multi-Exposure image Fusion Network(MEF-Net) based on deep guided and self-learning is proposed. This network is designed to fuse any number of images with different exposures in an end-to-end way, and generate the best-fused results in an unsupervised way. In terms of the loss function, an intensity fidelity constraint term and the weighted Multi-Exposure image Fusion Structural SIMilarity(MEF-SSIM) are introduced to improve the fusion quality. Moreover, a self-learning method is adopted to fine-tune and optimize the pre-learned model, considering the fusion problem of two images under extreme exposure to mitigate the halo phenomenon generated by fusion. Based on abundant testing data, experimental results show that the proposed algorithm outperforms other mainstream methods in terms of both quantitative measurement and visual fused quality.
2023, 45(1): 300-307.
doi: 10.11999/JEIT211199
Abstract:
The video frame type decision is one of the key factors affecting the efficiency of video coding. This paper proposes an adaptive video frame type decision algorithm based on local luminance histogram to improve the x265 encoding performance. Firstly, the local luminance histograms of frames are calculated at the level of 64×64 Coding Tree Unit (CTU), and the difference of local luminance histogram between frames is used to represent the degree of scene variation between frames. Secondly, Intra-coded picture (I-frame) detection window is introduced. I-frame is determined by comparing the degree of scene variation between frames. Finally, the Mini Group Of Picture (MiniGOP) size is determined according to the correlation between the degree of scene variation and MiniGOP size, so as to determine adaptively Generalized P and B picture (GPB-frame) and Bidirectionally predicted picture (B-frame). Experimental results show that compared with the relevant algorithms in x265, the proposed algorithm can effectively reduce the coding complexity of x265, and decide I/GPB/B-frame efficiently and adaptively with nearly 5% less coding time.
The video frame type decision is one of the key factors affecting the efficiency of video coding. This paper proposes an adaptive video frame type decision algorithm based on local luminance histogram to improve the x265 encoding performance. Firstly, the local luminance histograms of frames are calculated at the level of 64×64 Coding Tree Unit (CTU), and the difference of local luminance histogram between frames is used to represent the degree of scene variation between frames. Secondly, Intra-coded picture (I-frame) detection window is introduced. I-frame is determined by comparing the degree of scene variation between frames. Finally, the Mini Group Of Picture (MiniGOP) size is determined according to the correlation between the degree of scene variation and MiniGOP size, so as to determine adaptively Generalized P and B picture (GPB-frame) and Bidirectionally predicted picture (B-frame). Experimental results show that compared with the relevant algorithms in x265, the proposed algorithm can effectively reduce the coding complexity of x265, and decide I/GPB/B-frame efficiently and adaptively with nearly 5% less coding time.
2023, 45(1): 308-316.
doi: 10.11999/JEIT211176
Abstract:
Considering the problem of low recognition accuracy of Specific Emitter Identification (SEI) and high cost of single training, an SEI scheme based on incremental learning is proposed in this paper, multiple Continuous Incremental Deep Extreme Learning Machine(CIDELM) are designed. The Hilbert spectrum projection and higher-order spectrum processed by Variational Mode Decomposition (VMD) are extracted from the original signal, and they are used as the Radio Fingerprint Feature (RFF) for classification after dimensionality reduction. In the Extreme Learning Machine (ELM), the sparse self-encoding structure is introduced to perform unsupervised training on multiple hidden layers, and the parameter search strategy is used to determine the best number of hidden layers and hidden nodes, realizing online multi-batch labeled samples continuous matching. The results show that the algorithm can show good compatibility with different modulation modes, carrier frequencies and transmission distances, and can effectively identify multiple transmitters.
Considering the problem of low recognition accuracy of Specific Emitter Identification (SEI) and high cost of single training, an SEI scheme based on incremental learning is proposed in this paper, multiple Continuous Incremental Deep Extreme Learning Machine(CIDELM) are designed. The Hilbert spectrum projection and higher-order spectrum processed by Variational Mode Decomposition (VMD) are extracted from the original signal, and they are used as the Radio Fingerprint Feature (RFF) for classification after dimensionality reduction. In the Extreme Learning Machine (ELM), the sparse self-encoding structure is introduced to perform unsupervised training on multiple hidden layers, and the parameter search strategy is used to determine the best number of hidden layers and hidden nodes, realizing online multi-batch labeled samples continuous matching. The results show that the algorithm can show good compatibility with different modulation modes, carrier frequencies and transmission distances, and can effectively identify multiple transmitters.
2023, 45(1): 317-324.
doi: 10.11999/JEIT211126
Abstract:
Gene expression data is usually characterized by high dimension, few samples and uneven classification distribution. How to extract the effective features of gene expression data is a critical problem of gene classification. With the help of correlation analysis theory, the within-view and between-view discrimination sensitive similarity order scatter can be construsted, thus forming a new method of gene feature extraction, namely, Similarity Order Preserving Across-view Correlation Analysis(SOPACA). The proposed method not only maintains the intra-class aggregation and similarity order of features between different views, but also has a large distance between classes. Good experimental results on cancer gene expression datasets demonstrate the effectiveness of the method.
Gene expression data is usually characterized by high dimension, few samples and uneven classification distribution. How to extract the effective features of gene expression data is a critical problem of gene classification. With the help of correlation analysis theory, the within-view and between-view discrimination sensitive similarity order scatter can be construsted, thus forming a new method of gene feature extraction, namely, Similarity Order Preserving Across-view Correlation Analysis(SOPACA). The proposed method not only maintains the intra-class aggregation and similarity order of features between different views, but also has a large distance between classes. Good experimental results on cancer gene expression datasets demonstrate the effectiveness of the method.
2023, 45(1): 325-334.
doi: 10.11999/JEIT211212
Abstract:
Cross-modal person Re-IDentification (Re-ID) is a challenging problem for intelligent surveillance systems, and existing cross-modal research approaches are mainly based on global or local learning representation of differentiated modal shared features. However, few studies have attempted fuse global and local feature representations. A new Multi-granularity Shared Feature Fusion (MSFF) network is proposed in this paper, which combines global and local features to learn different granularities representations of the two modalities, extracting multi-scale and multi-level features from the backbone network, where the coarse granularity information of the global feature representation and the fine granularity information of the local feature representation collaborate with each other to form more differentiated feature descriptors. In addition, in order to extract more effective shared features for the network, the paper also proposes an improved method of subspace shared feature module for embedding modes of the two modalities in the network, changing the feature embedding mode of traditional modal feature weights. The module is put into the backbone network in advance so that the respective features of the two modalities are mapped into the same subspace to generate richer shared weights through the backbone network. The experimental results in two public datasets demonstrate the effectiveness of the proposed method, and the average accuracy mAP in the most difficult full-search single-shot mode of SYSU-MM01 dataset reaches 60.62%.
Cross-modal person Re-IDentification (Re-ID) is a challenging problem for intelligent surveillance systems, and existing cross-modal research approaches are mainly based on global or local learning representation of differentiated modal shared features. However, few studies have attempted fuse global and local feature representations. A new Multi-granularity Shared Feature Fusion (MSFF) network is proposed in this paper, which combines global and local features to learn different granularities representations of the two modalities, extracting multi-scale and multi-level features from the backbone network, where the coarse granularity information of the global feature representation and the fine granularity information of the local feature representation collaborate with each other to form more differentiated feature descriptors. In addition, in order to extract more effective shared features for the network, the paper also proposes an improved method of subspace shared feature module for embedding modes of the two modalities in the network, changing the feature embedding mode of traditional modal feature weights. The module is put into the backbone network in advance so that the respective features of the two modalities are mapped into the same subspace to generate richer shared weights through the backbone network. The experimental results in two public datasets demonstrate the effectiveness of the proposed method, and the average accuracy mAP in the most difficult full-search single-shot mode of SYSU-MM01 dataset reaches 60.62%.
2023, 45(1): 335-343.
doi: 10.11999/JEIT211288
Abstract:
Occluded object segmentation is a difficult problem in instance segmentation, but it has great practical value in many industrial applications such as stacked parcel segmentation on logistics automatic sorting. In this paper, an occluded object segmentation method based on bilayer decoupling strategy and attention mechanism is proposed to improve the segmentation performance of occluded parcels. Firstly, the image features are extracted through a backbone network with a Feature Pyramid Network (FPN); Secondly, the bilayer decoupling head is used to predict whether the mass centers of instances are occluded, and different occlusion types of instances are predicted through different branches; Thirdly, attention refinement module is used to obtain predicted masks of non-occluded instances and generate an attention map by combining these masks; Finally, this attention map is used to help the prediction of occluded instances. A dataset is provided for occluded parcel segmentation. Our method is tested on this dataset. The experimental results show that the proposed network achieves 95,66% Average Precision(AP), 97.17% Recall, and 11.78% Miss Rate(MR–2). It indicates that this method has better segmentation performance than other methods.
Occluded object segmentation is a difficult problem in instance segmentation, but it has great practical value in many industrial applications such as stacked parcel segmentation on logistics automatic sorting. In this paper, an occluded object segmentation method based on bilayer decoupling strategy and attention mechanism is proposed to improve the segmentation performance of occluded parcels. Firstly, the image features are extracted through a backbone network with a Feature Pyramid Network (FPN); Secondly, the bilayer decoupling head is used to predict whether the mass centers of instances are occluded, and different occlusion types of instances are predicted through different branches; Thirdly, attention refinement module is used to obtain predicted masks of non-occluded instances and generate an attention map by combining these masks; Finally, this attention map is used to help the prediction of occluded instances. A dataset is provided for occluded parcel segmentation. Our method is tested on this dataset. The experimental results show that the proposed network achieves 95,66% Average Precision(AP), 97.17% Recall, and 11.78% Miss Rate(MR–2). It indicates that this method has better segmentation performance than other methods.
2023, 45(1): 344-352.
doi: 10.11999/JEIT211147
Abstract:
TweAES is one of the second-round candidates in the NIST Lightweight Cryptography Standardization competition. The related-tweak multiple impossible differentials attack of 8-round TweAES is presented. Firstly, two types of impossible differential distinguishers are utilized to construct two attack trails, and each attack trail needs to guess 16 Byte subkey. It is worth noting that two attack trails have the same plaintext structure and 14 Byte common subkey. Attackers can utilize the plaintext pairs with the same plaintext structure to reject wrong subkeys by two filters processed, and because of a large number of common subkey, the efficiency of subkeys sifting can be improved. Furthermore, the incompleteness of the key schedule is utilized to choose the subkey Bytes. With the help of the relations of subkey Bytes, the efficiency of reconstructing the corresponding master keys can be improved, so the complexity of the whole attack scheme can be improved. Compared with the previous results, this work obtain the new attack scheme of 8-round TweAES, which needs lower time, data, and memory complexities than other attack schemes.
TweAES is one of the second-round candidates in the NIST Lightweight Cryptography Standardization competition. The related-tweak multiple impossible differentials attack of 8-round TweAES is presented. Firstly, two types of impossible differential distinguishers are utilized to construct two attack trails, and each attack trail needs to guess 16 Byte subkey. It is worth noting that two attack trails have the same plaintext structure and 14 Byte common subkey. Attackers can utilize the plaintext pairs with the same plaintext structure to reject wrong subkeys by two filters processed, and because of a large number of common subkey, the efficiency of subkeys sifting can be improved. Furthermore, the incompleteness of the key schedule is utilized to choose the subkey Bytes. With the help of the relations of subkey Bytes, the efficiency of reconstructing the corresponding master keys can be improved, so the complexity of the whole attack scheme can be improved. Compared with the previous results, this work obtain the new attack scheme of 8-round TweAES, which needs lower time, data, and memory complexities than other attack schemes.
2023, 45(1): 353-360.
doi: 10.11999/JEIT211235
Abstract:
Linear Complementary Dual (LCD) codes have important applications to resisting side-channel analysis and fault-injection attacks. A method is proposed to construct ternary LCD codes by using linear codes over ring\begin{document}$ {\mathbb{F}_3} + u{\mathbb{F}_3} $\end{document} ![]()
![]()
, where \begin{document}$ {u^2} = 0 $\end{document} ![]()
![]()
. A Gray map from \begin{document}$ {({\mathbb{F}_3} + u{\mathbb{F}_3})^n} $\end{document} ![]()
![]()
to \begin{document}$ \mathbb{F}_3^{2n} $\end{document} ![]()
![]()
is introduced, and a sufficient condition for the Gray image of linear codes with length \begin{document}$ n $\end{document} ![]()
![]()
over \begin{document}$ {\mathbb{F}_3} + u{\mathbb{F}_3} $\end{document} ![]()
![]()
to be ternary LCD codes with length \begin{document}$ 2n $\end{document} ![]()
![]()
is given. Four classes of ternary LCD codes with better parameters are constructed via the Gray image of cyclic codes over \begin{document}$ {\mathbb{F}_3} + u{\mathbb{F}_3} $\end{document} ![]()
![]()
.
Linear Complementary Dual (LCD) codes have important applications to resisting side-channel analysis and fault-injection attacks. A method is proposed to construct ternary LCD codes by using linear codes over ring
2023, 45(1): 361-370.
doi: 10.11999/JEIT211297
Abstract:
Among the Compressed Sensing(CS) reconstruction algorithms, greedy algorithm has been widely studied for its simple hardware implementation and excellent recovery accuracy. However, the diversity of algorithms also presents the problem of difficult algorithm selection. As the core of greedy algorithms, the difference of atomic recognition strategy often determines its recovery performance. In this paper, atomic recognition strategy, which is the most important part of greedy algorithm, is taken as the research object. Three one-step atom recognition strategies, eight advanced atom recognition strategies and three sparsity adaptive atom recognition strategies are summarized according to the applicable stages and characteristics. Finally, the recovery performance of the original algorithms corresponding to the atomic recognition strategies are simulated and compared. The sorted strategies are convenient for the selection of algorithms in practical application, and they provide references for the further optimization of greedy algorithms.
Among the Compressed Sensing(CS) reconstruction algorithms, greedy algorithm has been widely studied for its simple hardware implementation and excellent recovery accuracy. However, the diversity of algorithms also presents the problem of difficult algorithm selection. As the core of greedy algorithms, the difference of atomic recognition strategy often determines its recovery performance. In this paper, atomic recognition strategy, which is the most important part of greedy algorithm, is taken as the research object. Three one-step atom recognition strategies, eight advanced atom recognition strategies and three sparsity adaptive atom recognition strategies are summarized according to the applicable stages and characteristics. Finally, the recovery performance of the original algorithms corresponding to the atomic recognition strategies are simulated and compared. The sorted strategies are convenient for the selection of algorithms in practical application, and they provide references for the further optimization of greedy algorithms.
2023, 45(1): 371-382.
doi: 10.11999/JEIT211185
Abstract:
In the cloud era, cloud Application Programming Interface (API) is the best carrier for service delivery, capability replication and data output. However, cloud API increases the exposure and attack surface of cloud application while opening up services and data. Through data hijacking, traffic analysis and other technologies, attackers can obtain the key resources of the target cloud API, so as to identify the identity and behavior of users, or even directly cause the paralysis of the underlying system. Currently, there are many types of attacks against cloud APIs, and their threats and protection methods are different. However, the existing researches lack a systematic summary for cloud API attack and protection methods. In this paper, a detail survey on the threats and protection methods faced by cloud API is conducted. Firstly, the evolution and the classification of cloud API are analyzed. The vulnerability of cloud API and the importance of cloud API security research are then discussed. Furthermore, a systematical cloud API security research framework is proposed, which covers six aspects: identity authentication, cloud API Distributed Denial of Service (DDoS) attack protection, replay attack protection, Man-In-The-Middle (MITM) attack protection, injection attack protection and sensitive data protection. In addition, the necessity of Artificial Intelligence (AI) protection for cloud API is discussed. Finally, the future challenges and development trends of cloud API protection are presented.
In the cloud era, cloud Application Programming Interface (API) is the best carrier for service delivery, capability replication and data output. However, cloud API increases the exposure and attack surface of cloud application while opening up services and data. Through data hijacking, traffic analysis and other technologies, attackers can obtain the key resources of the target cloud API, so as to identify the identity and behavior of users, or even directly cause the paralysis of the underlying system. Currently, there are many types of attacks against cloud APIs, and their threats and protection methods are different. However, the existing researches lack a systematic summary for cloud API attack and protection methods. In this paper, a detail survey on the threats and protection methods faced by cloud API is conducted. Firstly, the evolution and the classification of cloud API are analyzed. The vulnerability of cloud API and the importance of cloud API security research are then discussed. Furthermore, a systematical cloud API security research framework is proposed, which covers six aspects: identity authentication, cloud API Distributed Denial of Service (DDoS) attack protection, replay attack protection, Man-In-The-Middle (MITM) attack protection, injection attack protection and sensitive data protection. In addition, the necessity of Artificial Intelligence (AI) protection for cloud API is discussed. Finally, the future challenges and development trends of cloud API protection are presented.