Multi-agent privacy protection analysis method and system of pulse communication mechanism

Abstract

The invention relates to the technical field of computer security, in particular to a multi-agent privacy protection analysis method and system of a pulse communication mechanism. The method comprises the following steps: acquiring a continuous input signal sequence corresponding to a multi-agent system, and performing space-time sparse pulse conversion to generate a multi-agent space-time sparse pulse sequence; a layered space-time pulse neural network is constructed, pulse communication decision output and pulse disturbance randomization processing based on combination of pulse time jitter and pulse deletion are carried out in combination with a pulse network training mode based on differential privacy, and a multi-agent pulse disturbance randomization sequence is generated; and constructing privacy leakage attack scene models corresponding to the pulse sequence reconstruction attack, the member reasoning attack and the attribute reasoning attack, and performing privacy leakage attack simulation and privacy leakage protection analysis to generate a privacy-efficiency protection optimization strategy corresponding to multi-agent pulse communication. According to the invention, balance optimization of privacy protection and communication efficiency in a multi-agent system can be realized.

Description

Technical Field

[0001] The present invention relates to the field of computer security technology, and in particular to a multi-agent privacy protection analysis method and system for a pulse communication mechanism. Background Art

[0002] With the rapid development of intelligent technology, multi-agent systems (MAS) have been widely applied in many fields, such as intelligent transportation, intelligent manufacturing, robotic collaboration, and distributed energy management. These MAS systems achieve efficient task completion through mutual collaboration and information sharing. The pulse communication mechanism (Pulse Communication Mechanism) is a communication method that mimics the signal transmission of biological nervous systems. In this mechanism, information is transmitted through the triggering and propagation of pulse signals, rather than traditional packet exchange. This communication mechanism is highly time-efficient and has low latency, making it suitable for multi-agent systems that require fast response and high-frequency data exchange. However, traditional privacy protection methods currently rely mainly on encryption, homomorphic encryption, and private computing. However, these methods still face a key challenge when dealing with the high-frequency, real-time nature of pulse communication mechanisms: how to ensure efficient information transmission while protecting the privacy of agents. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide a multi-agent privacy protection analysis method and system for a pulse communication mechanism to solve at least one of the above technical problems.

[0004] To achieve the above objectives, a multi-agent privacy protection analysis method for an impulse communication mechanism is proposed, comprising the following steps: Step S1: obtaining a continuous input signal sequence corresponding to the multi-agent system, and performing spatiotemporal sparse pulse conversion on the continuous input signal sequence corresponding to the multi-agent system based on a sparse matrix generation method of time coding and rate coding to generate a multi-agent spatiotemporal sparse pulse sequence; Step S2: Construct a hierarchical spatiotemporal pulse neural network, and input the multi-agent spatiotemporal sparse pulse sequence into the hierarchical spatiotemporal pulse neural network. Combined with the pulse network training method based on differential privacy, the pulse communication decision output is performed to generate the multi-agent time-critical communication pulse sequence. Step S3: performing pulse perturbation randomization processing based on a combination of pulse time jitter and pulse deletion on the multi-agent timing-critical communication pulse sequence to generate a multi-agent pulse perturbation randomized sequence; Step S4: Construct privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, member inference attack and attribute inference attack, and use Monte Carlo simulation combined with the corresponding privacy leakage attack scenario model to simulate the privacy leakage attack on the multi-agent pulse perturbation randomized sequence to generate the corresponding multi-agent pulse privacy leakage probability under different attack scenarios; based on the corresponding multi-agent pulse privacy leakage probability under different attack scenarios, perform privacy leakage protection analysis on the multi-agent pulse perturbation randomized sequence to generate a privacy-efficiency protection optimization strategy corresponding to the multi-agent pulse communication.

[0005] Furthermore, step S1 includes the following steps: Step S11: Obtain a continuous input signal sequence corresponding to the multi-agent system; Step S12: performing input signal pulse generation on the continuous input signal sequence corresponding to the multi-agent system to generate a multi-agent input signal pulse sequence; Step S13: performing a time-delay sparse matrix conversion on the multi-agent input signal pulse sequence based on a sparse matrix generation method of time coding to generate a multi-agent signal sparse matrix with time delay characteristics; Step S14: performing pulse rate sparse matrix conversion on the multi-agent input signal pulse sequence based on the rate coding sparse matrix generation method to generate a multi-agent signal sparse matrix with rate characteristics; Step S15: Based on the multi-agent signal sparse matrix with time delay characteristics and rate characteristics, the multi-agent input signal pulse sequence is subjected to spatiotemporal sparse pulse conversion to generate a multi-agent spatiotemporal sparse pulse sequence.

[0006] Furthermore, step S13 includes the following steps: Step S131: obtaining corresponding signal pulse time intervals and signal pulse waveform characteristics through a multi-agent input signal pulse sequence; Step S132: Determine the thinning time window based on the signal pulse time interval and the signal pulse waveform characteristics, and obtain the signal pulse thinning time window size related to the time distribution of the original multi-agent signal; Step S133: Based on the signal pulse sparse time window size related to the time distribution of the multi-agent original signal, the pulse amplitude and time series position statistics are performed for each sparse time window moment in the multi-agent input signal pulse sequence to obtain the multi-agent signal pulse amplitude and multi-agent signal time series position corresponding to each sparse time window moment; Step S134: Calculate the delay distribution difference at different moments based on the multi-agent signal pulse amplitude and the multi-agent signal timing position corresponding to each sparse time window moment, and obtain the delay distribution difference between the multi-agent signal pulses at different sparse time windows; Step S135: Based on the delay distribution differences between multi-agent signal pulses at different sparse time windows and combined with the time-coded sparse matrix generation method, the multi-agent input signal pulse sequence is transformed into a delay sparse matrix to generate a multi-agent signal sparse matrix with delay characteristics.

[0007] Furthermore, step S14 includes the following steps: Step S141: performing statistical calculations on the pulse frequency and amplitude of each signal pulse window of the multi-agent input signal pulse sequence to obtain the corresponding multi-agent signal pulse frequency and multi-agent signal pulse amplitude in each signal pulse window; Step S142: obtaining a corresponding multi-agent pulse signal strength through the multi-agent input signal pulse sequence, and performing pulse rate modulation on the multi-agent input signal pulse sequence based on the multi-agent pulse signal strength to generate a modulated signal pulse rate corresponding to the multi-agent signal pulse; Step S143: Based on the modulation signal pulse rate corresponding to the multi-agent signal pulse and combined with the rate coding sparse matrix generation method, the multi-agent signal pulse frequency and the multi-agent signal pulse amplitude corresponding to each signal pulse window are sparsely weighted reconstructed to generate a multi-agent signal sparse matrix with rate characteristics.

[0008] Furthermore, step S15 includes the following steps: Step S151: performing time-delay sparse filtering on the multi-agent input signal pulse sequence based on the multi-agent signal sparse matrix with time-delay characteristics to obtain a multi-agent time-delay sparse filtered pulse sequence; Step S152: performing rate sparse modulation on the multi-agent time-delay sparse filtered pulse sequence in spatial position distribution based on the multi-agent signal sparse matrix with rate characteristics to generate a multi-agent rate sparse modulated pulse sequence; Step S153: Obtain the pulse energy corresponding to each pulse signal through the multi-agent rate sparse modulation pulse sequence, and perform pulse attenuation evaluation based on the pulse energy corresponding to each pulse signal to obtain the pulse attenuation factor corresponding to the multi-agent pulse signal; Step S154: Based on the pulse attenuation factor corresponding to the multi-agent pulse signal, the spatiotemporal pulse distribution of each corresponding pulse signal in the multi-agent rate sparse modulation pulse sequence is optimized to generate a multi-agent spatiotemporal sparse pulse sequence.

[0009] Furthermore, step S2 includes the following steps: Step S21: Constructing a corresponding hierarchical spatiotemporal spike neural network by adopting a structure combining a recursive neural network and a long short-term memory network, specifically including a bottom layer, a middle layer, and a top layer, wherein the bottom layer is used to process the original pulse sequence, the middle layer is used to extract the corresponding pulse timing features, and the top layer is used to decide and output the corresponding communication pulse sequence; Step S22: Input the multi-agent spatiotemporal sparse pulse sequence into the bottom layer of the hierarchical spatiotemporal spiking neural network for pulse preprocessing, and input the preprocessed multi-agent spatiotemporal sparse pulse sequence into the middle layer of the hierarchical spatiotemporal spiking neural network to train the pulse extraction network by introducing a pulse timing attention mechanism. At the same time, a differentially private spike network training method is designed and constructed during the training process to introduce Gaussian noise into the spike neural network and combine the gradient clipping strategy of the training process to prevent gradient reversal attacks, so as to train and generate the pulse timing feature output corresponding to each agent; Step S23: The pulse interaction frequency and pulse sensitivity between each agent are calculated by outputting the pulse timing characteristics corresponding to each agent at the top layer of the hierarchical spatiotemporal pulse neural network, and the corresponding privacy budget is dynamically configured according to the pulse interaction frequency and pulse sensitivity between each agent. parameters, and according to the privacy budget The parameters in the top layer decide the corresponding communication pulse sequence output to generate the multi-agent timing-critical communication pulse sequence.

[0010] Furthermore, step S3 includes the following steps: Step S31: performing pulse timing disturbance analysis on the multi-agent timing-critical communication pulse sequence to obtain the multi-agent communication pulse timing disturbance distribution characteristics; Step S32: Obtain the pulse communication disturbance delay and pulse communication disturbance frequency difference between each agent through the multi-agent communication pulse timing disturbance distribution characteristics, and perform pulse time jitter amplitude statistics on each pulse signal in the multi-agent timing critical communication pulse sequence based on the pulse communication disturbance delay and pulse communication disturbance frequency difference between each agent to obtain the communication pulse signal time jitter amplitude between each agent; Step S33: Generate a corresponding multi-agent pulse time jitter amplitude spectrum based on the time jitter amplitude of the communication pulse signals between each agent, and perform pulse selective deletion optimization on each pulse signal in the multi-agent timing-critical communication pulse sequence based on the multi-agent pulse time jitter amplitude spectrum to filter out redundant pulse signals corresponding to the consistent time jitter amplitude of the communication pulse signals, thereby generating a multi-agent pulse optimized deletion sequence; Step S34: Perform pulse perturbation distribution randomization processing on the timing of the corresponding remaining pulse signals in the multi-agent pulse optimization deletion sequence to generate a multi-agent pulse perturbation randomization sequence.

[0011] Furthermore, step S4 includes the following steps: Step S41: constructing privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, membership inference attack, and attribute inference attack; Step S42: Monte Carlo simulation is used in combination with the corresponding privacy leakage attack scenario model to simulate the privacy leakage attack on the randomized sequence of multi-agent pulse perturbations to generate the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios; Step S43: Based on the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios, the privacy leakage diffusion impact analysis is performed on the multi-agent pulse perturbation randomized sequence to obtain the corresponding diffusion path and impact range of the corresponding pulse privacy leakage event in the multi-agent system; Step S44: Constructing a privacy-efficiency optimization target corresponding to the privacy protection level, communication bandwidth consumption, and computational delay based on the diffusion path and impact range of the corresponding pulse privacy leakage event in the multi-agent system; Step S45: Based on the privacy-efficiency optimization objectives corresponding to the privacy protection level, communication bandwidth consumption and computing delay, a privacy leakage protection analysis is performed on the randomized sequence of multi-agent pulse perturbations, and the Pareto optimal solution set corresponding to the privacy-efficiency optimization objectives is solved by the NSGA-II algorithm. Based on the Pareto optimal solution set, a balance strategy between privacy and efficiency corresponding to the pulse signal communication is obtained to generate a privacy-efficiency protection optimization strategy corresponding to the multi-agent pulse communication.

[0012] Furthermore, step S42 includes the following steps: Monte Carlo simulation is used to combine the privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, membership inference attack and attribute inference attack to simulate and infer the pulse attack of the randomized multi-agent pulse perturbation sequence, so as to generate the corresponding multi-agent pulse perturbation attack generation process under different attack scenarios. Perform multiple rounds of reverse reasoning on the pulse decision-making process of the multi-agent pulse perturbation attack generation process under different attack scenarios to generate the multi-agent pulse perturbation attack decision tree corresponding to different attack scenarios; The corresponding multi-agent pulse perturbation stagnation points in different attack scenarios are obtained through the multi-agent pulse perturbation attack decision trees in different attack scenarios, and the corresponding multi-agent pulse perturbation stagnation points in different attack scenarios are used to predict the leakage probability of the multi-agent pulse perturbation attack decision trees in different attack scenarios to generate the corresponding multi-agent pulse privacy leakage probability in different attack scenarios.

[0013] Furthermore, the present invention also provides a multi-agent privacy protection analysis system for an impulse communication mechanism, which is used to perform the multi-agent privacy protection analysis method for the impulse communication mechanism described above. The multi-agent privacy protection analysis system for the impulse communication mechanism includes: A spatiotemporal sparse pulse conversion module is used to obtain a continuous input signal sequence corresponding to the multi-agent system and perform spatiotemporal sparse pulse conversion on the continuous input signal sequence corresponding to the multi-agent system based on a sparse matrix generation method of time coding and rate coding to generate a multi-agent spatiotemporal sparse pulse sequence; The pulse communication decision output module is used to construct a hierarchical spatiotemporal pulse neural network and input the multi-agent spatiotemporal sparse pulse sequence into the hierarchical spatiotemporal pulse neural network. Combined with the pulse network training method based on differential privacy, the pulse communication decision output is performed to generate the multi-agent time-critical communication pulse sequence; The pulse perturbation randomization module is used to perform pulse perturbation randomization processing on the multi-agent timing-critical communication pulse sequence based on the combination of pulse time jitter and pulse deletion, thereby generating a multi-agent pulse perturbation randomized sequence; The pulse privacy protection analysis module is used to construct privacy leakage attack scenario models corresponding to pulse sequence reconstruction attacks, member inference attacks, and attribute inference attacks, and uses Monte Carlo simulation combined with the corresponding privacy leakage attack scenario models to simulate privacy leakage attacks on multi-agent pulse perturbation randomized sequences to generate the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios; based on the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios, the privacy leakage protection analysis of the multi-agent pulse perturbation randomized sequences is performed to generate the privacy-efficiency protection optimization strategy corresponding to multi-agent pulse communication.

[0014] Beneficial effects of the present invention: 1. The multi-agent privacy protection analysis method of the pulse communication mechanism proposed in the present invention has the beneficial effect of obtaining a continuous input signal sequence corresponding to the multi-agent system and converting these signals into spatiotemporal sparse pulses through a sparse matrix generation method based on time coding and rate coding. The core advantage of this step is that the continuous input signal is effectively converted into a sparse pulse sequence through spatiotemporal sparse coding, which not only reduces the amount of data transmission but also improves the efficiency of information processing. The sparse coding method can effectively compress the signal and reduce the computational burden of the system while ensuring that key information is not lost. Especially in a multi-agent system, the number of agents is very large and the complexity of the signal increases accordingly. Through this spatiotemporal sparse pulse conversion, unnecessary computational overhead can be greatly reduced, and the communication efficiency between multiple agents can be optimized. The generation of spatiotemporal sparse pulse sequences also provides more efficient and concise data input for subsequent pulse neural networks, laying a solid foundation for improving the overall performance of the system. Secondly, by constructing a hierarchical spatiotemporal spike neural network, the signal sequence that has undergone spatiotemporal sparse spike conversion is input into the network. Combined with a spike network training method based on differential privacy, pulse communication decision output is performed, thereby generating a multi-agent time-critical communication pulse sequence. This fully exploits the spatiotemporal characteristics of the input signal and decomposes and processes the complex input signal at different levels, thereby improving the overall processing power and accuracy of the multi-agent system. Through the hierarchical design, the neural network can capture the key characteristics of the signal at multiple scales, ensuring more efficient and accurate information transmission in the multi-agent system. The introduction of differential privacy effectively enhances data protection throughout the spike network training process, avoiding the risk of data leakage. Privacy protection is a non-negligible issue in multi-agent systems, and differential privacy technology provides a viable solution. By encrypting the training process, normal operation and decision output can still be achieved while ensuring data privacy. This training method combined with privacy protection provides higher security and reliability for the application of multi-agent systems.Then, the multi-agent timing-critical communication pulse sequence is subjected to pulse perturbation randomization processing based on a combination of pulse time jitter and pulse deletion. The main advantage of this process is that by introducing perturbation randomization, the privacy protection effect can be further improved while maintaining the efficiency and stability of the system. Pulse time jitter and pulse deletion are two effective privacy protection methods. They can disrupt the attacker's analysis of data without significantly affecting the system performance, thereby effectively preventing potential attacks. Pulse time jitter randomly changes the time interval of the pulses, making it difficult for the attacker to accurately infer the timing relationship of the signal, while pulse deletion deletes part of the pulse information, further increasing the difficulty for the attacker to obtain the complete signal. This can ensure the privacy of the agents while ensuring the efficiency of information transmission, thereby ensuring the privacy security and functional stability of the multi-agent system. Finally, by constructing privacy leakage attack scenario models of pulse sequence reconstruction attack, member inference attack and attribute inference attack, and combining Monte Carlo simulation to simulate privacy leakage attack on multi-agent pulse perturbation randomized sequence, the core advantage of this process is that by simulating multiple attack scenarios, the privacy protection ability of the system under different attack conditions can be comprehensively evaluated. Through Monte Carlo simulation, the robustness of the system can be verified in a large number of experiments, and potential weaknesses can be found. Privacy leakage attack simulation not only helps designers understand the effectiveness of current protection measures, but also provides an important reference for subsequent optimization. Based on the simulation, the privacy leakage protection analysis of the pulse perturbation randomized sequence can be carried out in a targeted manner, and finally the optimization strategy is generated. These strategies will weigh the relationship between privacy protection and system efficiency, and provide the best privacy protection and performance balance solution. This privacy-efficiency protection optimization strategy can ensure that the multi-agent system can still maintain a high degree of privacy security in the face of complex attacks, while ensuring the efficiency of the system in practical applications.

[0015] 2. The multi-agent privacy protection analysis system for the pulse communication mechanism proposed in the present invention is generally composed of a spatiotemporal sparse pulse conversion module, a pulse communication decision output module, a pulse perturbation randomization module, and a pulse privacy protection analysis module. It can implement the multi-agent privacy protection analysis method for any pulse communication mechanism described in the present invention, and is used to combine the operations between the computer programs running on each module to implement the multi-agent privacy protection analysis method for the pulse communication mechanism. The internal structures of the system cooperate with each other, which can greatly reduce duplication of work and manpower investment, and can quickly and effectively provide a more accurate and efficient multi-agent privacy protection analysis process for the pulse communication mechanism, thereby simplifying the operation process of the multi-agent privacy protection analysis system for the pulse communication mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings: Figure 1 Schematic diagram of the steps of the multi-agent privacy protection analysis method of the pulse communication mechanism of the present invention; Figure 2 for Figure 1 Detailed step flow diagram of step S1; Figure 3 for Figure 2 Detailed step flow chart of step S13 in FIG. DETAILED DESCRIPTION

[0017] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0018] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0019] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0020] To achieve this, please refer to Figures 1 to 3 The present invention provides a multi-agent privacy protection analysis method for an impulse communication mechanism, the method comprising the following steps: Step S1: obtaining a continuous input signal sequence corresponding to the multi-agent system, and performing spatiotemporal sparse pulse conversion on the continuous input signal sequence corresponding to the multi-agent system based on a sparse matrix generation method of time coding and rate coding to generate a multi-agent spatiotemporal sparse pulse sequence; Step S2: Construct a hierarchical spatiotemporal pulse neural network, and input the multi-agent spatiotemporal sparse pulse sequence into the hierarchical spatiotemporal pulse neural network. Combined with the pulse network training method based on differential privacy, the pulse communication decision output is performed to generate the multi-agent time-critical communication pulse sequence. Step S3: performing pulse perturbation randomization processing based on a combination of pulse time jitter and pulse deletion on the multi-agent timing-critical communication pulse sequence to generate a multi-agent pulse perturbation randomized sequence; Step S4: Construct privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, member inference attack and attribute inference attack, and use Monte Carlo simulation combined with the corresponding privacy leakage attack scenario model to simulate the privacy leakage attack on the multi-agent pulse perturbation randomized sequence to generate the corresponding multi-agent pulse privacy leakage probability under different attack scenarios; based on the corresponding multi-agent pulse privacy leakage probability under different attack scenarios, perform privacy leakage protection analysis on the multi-agent pulse perturbation randomized sequence to generate a privacy-efficiency protection optimization strategy corresponding to the multi-agent pulse communication.

[0021] In the embodiment of the present invention, please refer to Figure 1 FIG. 2 is a flow chart of the steps of the multi-agent privacy protection analysis method for the pulse communication mechanism of the present invention. In this example, the multi-agent privacy protection analysis method for the pulse communication mechanism includes the following steps: Step S1: obtaining a continuous input signal sequence corresponding to the multi-agent system, and performing spatiotemporal sparse pulse conversion on the continuous input signal sequence corresponding to the multi-agent system based on a sparse matrix generation method of time coding and rate coding to generate a multi-agent spatiotemporal sparse pulse sequence; In an embodiment of the present invention, in an industrial Internet of Things multi-agent system composed of 6 agents, data of each agent is collected in real time through a sensor network to obtain a continuous input signal sequence corresponding to the multi-agent system. Taking agent A as an example, the vibration sensor it carries collects equipment vibration data at a sampling frequency of 2000 Hz, and obtains 2000 continuous vibration amplitude data within 1 second. For example, the amplitudes of the 1st to 5th sampling points are 12mV, 15mV, 13mV, 14mV, and 16mV respectively. The other 5 agents also use their own sensors to collect temperature, pressure and other data at the same or different frequencies, which together constitute a continuous input signal sequence of the multi-agent system. The sparse matrix generation method based on time coding and rate coding is used to perform spatiotemporal sparse pulse conversion. In terms of time coding, the time axis is divided into a time window of 100 milliseconds. For the vibration signal sequence of agent A, if the time is between 200 and 300 milliseconds, In the window, the maximum signal amplitude is 20mV, which exceeds the set threshold of 18mV. It is determined that there is a pulse in the time window. In terms of rate coding, the number of pulses generated by each agent in 1 second is counted as the pulse rate. Agent A generated 8 pulses in the above 1 second, with a pulse rate of 8 times / second. A spatiotemporal sparse matrix with 6 rows (corresponding to 6 agents) and 10 columns (corresponding to 10 time windows) is constructed. For each agent in each time window, if there is a pulse, the pulse-related parameters (such as the normalized values ​​of the amplitude and rate) are filled in. If there is no pulse, 0 is filled in. For example, if agent A has a pulse in the third time window, its amplitude is normalized to 0.8 and its rate is normalized to 0.7. The corresponding position in the matrix is ​​filled with a vector composed of these two values ​​[0.8, 0.7]. By processing the continuous input signal sequence of all agents, a multi-agent spatiotemporal sparse pulse sequence is finally generated.

[0022] Step S2: Construct a hierarchical spatiotemporal pulse neural network, and input the multi-agent spatiotemporal sparse pulse sequence into the hierarchical spatiotemporal pulse neural network. Combined with the pulse network training method based on differential privacy, the pulse communication decision output is performed to generate the multi-agent time-critical communication pulse sequence. In an embodiment of the present invention, a hierarchical spatiotemporal pulse neural network is constructed, which includes a bottom layer, a middle layer and a top layer. The bottom layer is composed of 12 neurons, each neuron corresponds to a characteristic dimension of the agent signal, and receives a multi-agent spatiotemporal sparse pulse sequence as input; the middle layer is provided with 8 LSTM neurons for extracting pulse timing features; the top layer has 4 neurons, which are responsible for outputting pulse communication decisions, and inputting the multi-agent spatiotemporal sparse pulse sequence into the hierarchical spatiotemporal pulse neural network. At the bottom layer, the neurons perform weighted summation processing on the input pulses, and the weights are predetermined by the least squares method based on historical data. For example, for a pulse input of agent B, the weight of neuron 1 is 0.3, the pulse value is 0.6, and the weighted value is 0.3. 0.6=0.18, the processed signal is transmitted to the middle layer, and the LSTM neurons in the middle layer introduce the pulse timing attention mechanism. The attention weight calculation formula is ,in , is the hidden state at the current moment, For the The characteristic vector of the pulse, 、 、 is a trainable weight matrix. During the training process, a pulse network training method based on differential privacy is adopted, and Gaussian noise is added to the gradient of each training. The standard deviation of the noise is ,in is the sensitivity of the gradient, = 0.5 is the privacy budget parameter. A gradient clipping strategy is also used to clip the gradient norm to C = 0.8. After 200 training cycles, the network converges and ultimately outputs a multi-agent timing-critical communication pulse sequence at the top-level decision-making level.

[0023] Step S3: performing pulse perturbation randomization processing based on a combination of pulse time jitter and pulse deletion on the multi-agent timing-critical communication pulse sequence to generate a multi-agent pulse perturbation randomized sequence; In this embodiment of the present invention, a pulse perturbation randomization process based on a combination of pulse time jitter and pulse deletion is performed on a multi-agent timing-critical communication pulse sequence. The pulse time jitter range is set to ±50 milliseconds, and the pulse deletion probability is set to 0.2. For example, a pulse from agent C is originally generated at 0.5 seconds. A time offset is randomly generated within the jitter range. Assuming it is 20 milliseconds, the new generation time of the pulse becomes 0.5 + 0.02 = 0.52 seconds. For each pulse, a random number between 0 and 1 is generated and compared with the pulse deletion probability. If the random number is less than 0.2, the pulse is deleted. For example, if the random number generated is 0.15, which is less than 0.2, the pulse is deleted. This process is repeated for all pulses in the multi-agent timing-critical communication pulse sequence. For example, if the original sequence contains 100 pulses, 20 pulses are deleted after processing, and the remaining pulses are time-jittered, ultimately generating a multi-agent pulse perturbation randomization sequence.

[0024] Step S4: Construct privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, member inference attack and attribute inference attack, and use Monte Carlo simulation combined with the corresponding privacy leakage attack scenario model to simulate the privacy leakage attack on the multi-agent pulse perturbation randomized sequence to generate the corresponding multi-agent pulse privacy leakage probability under different attack scenarios; based on the corresponding multi-agent pulse privacy leakage probability under different attack scenarios, perform privacy leakage protection analysis on the multi-agent pulse perturbation randomized sequence to generate a privacy-efficiency protection optimization strategy corresponding to the multi-agent pulse communication.

[0025] In an embodiment of the present invention, by constructing privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, member inference attack and attribute inference attack, in the pulse sequence reconstruction attack scenario model, the attacker uses the known pulse generation rules and communication protocol, and adopts a reconstruction algorithm based on the least squares method to try to restore the original pulse sequence from the multi-agent pulse perturbation randomized sequence. Under the member inference attack scenario model, the attacker uses the support vector machine algorithm to determine whether a certain agent belongs to the system based on the historical communication data and the current observation sequence of the multi-agent system. In the attribute inference attack scenario model, the attacker constructs a random forest classification model based on the statistical characteristics of the agent pulse sequence to infer the agent attributes. Monte Carlo simulation is used in combination with the attack scenario model to simulate the privacy leakage attack on the multi-agent pulse perturbation randomized sequence, and the number of simulations is set to 10,000 times. In the pulse sequence reconstruction attack simulation, each simulation calculates the mean square error between the reconstructed sequence and the original sequence ,in is the original pulse value, is the reconstructed pulse value, if If the threshold exceeds 0.1, it is considered a privacy leak. After 10,000 simulations, the number of leaks is counted and the privacy leak probability is calculated. Assuming it is 0.18, the same simulation calculations are performed for member reasoning attacks and attribute reasoning attacks, resulting in privacy leak probabilities of 0.12 and 0.15, respectively. Based on these probabilities, the privacy leak diffusion path and impact range are analyzed, and a privacy-efficiency optimization objective is constructed with the privacy protection level (the inverse of the weighted sum of each attack probability), communication bandwidth consumption (calculated based on the number of pulses and transmission time), and computing delay (calculated based on the time it takes for the system to process pulses) as targets. Finally, the NSGA-II algorithm is used to solve this optimization objective and obtain a Pareto optimal solution set. From this solution, a suitable solution is selected to generate a privacy-efficiency protection optimization strategy for multi-agent pulse communication.

[0026] Furthermore, step S1 includes the following steps: Step S11: Obtain a continuous input signal sequence corresponding to the multi-agent system; Step S12: performing input signal pulse generation on the continuous input signal sequence corresponding to the multi-agent system to generate a multi-agent input signal pulse sequence; Step S13: performing a time-delay sparse matrix conversion on the multi-agent input signal pulse sequence based on a sparse matrix generation method of time coding to generate a multi-agent signal sparse matrix with time delay characteristics; Step S14: performing pulse rate sparse matrix conversion on the multi-agent input signal pulse sequence based on the rate coding sparse matrix generation method to generate a multi-agent signal sparse matrix with rate characteristics; Step S15: Based on the multi-agent signal sparse matrix with time delay characteristics and rate characteristics, the multi-agent input signal pulse sequence is subjected to spatiotemporal sparse pulse conversion to generate a multi-agent spatiotemporal sparse pulse sequence.

[0027] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed step flow diagram of step S1 in FIG. 1 , in this embodiment, step S1 includes the following steps: Step S11: Obtain a continuous input signal sequence corresponding to the multi-agent system; In an embodiment of the present invention, in a multi-agent system composed of 8 agents, the status data of each agent is collected in real time through a sensor array to obtain a continuous input signal sequence corresponding to the multi-agent system. Taking agent A1 as an example, the voltage sensor it carries continuously collects its own working voltage data at a sampling frequency of 1000 Hz, and obtains 1000 consecutive voltage values ​​within 1 second. For example, the voltage values ​​at the 1st to 5th sampling points are 3.2V, 3.25V, 3.3V, 3.28V, and 3.32V, respectively, forming a continuous voltage signal sequence. The other 7 agents also use their respective sensors to collect different types of status data such as current and temperature at the same sampling frequency, together forming a continuous input signal sequence of the multi-agent system. These data are transmitted to the data processing center through a wired transmission line and a fixed communication protocol for subsequent processing.

[0028] Step S12: performing input signal pulse generation on the continuous input signal sequence corresponding to the multi-agent system to generate a multi-agent input signal pulse sequence; In an embodiment of the present invention, input signal pulses are generated from a previously acquired continuous input signal sequence of a multi-agent system. The pulse generation threshold for the voltage signal is set to 3.3V. Taking the voltage signal sequence of agent A1 as an example, when the voltage value at a sampling point exceeds 3.3V for the first time, it is determined to be the starting point of a pulse. When the voltage value falls below 3.3V again, it is determined to be the end point of the pulse. In the above 1-second signal sequence, the voltage value at the third sampling point is 3.3V, and the voltage value at the eighth sampling point drops to 3.29V. Therefore, a pulse is formed from the third to the eighth sampling points, which contains data from six sampling points. The same threshold judgment method is used to generate pulses for various continuous input signal sequences of all agents. For current signals, the threshold is set to 5A; for temperature signals, the threshold is set to 40°C. After processing, the current signal sequence of agent A2 generates three pulses, and the temperature signal sequence of agent A3 generates one pulse. Finally, a complete multi-agent input signal pulse sequence is generated. This sequence contains pulse information of different types of signals from each agent, providing basic data for subsequent analysis.

[0029] Step S13: performing a time-delay sparse matrix conversion on the multi-agent input signal pulse sequence based on a sparse matrix generation method of time coding to generate a multi-agent signal sparse matrix with time delay characteristics; In an embodiment of the present invention, a time-delay sparse matrix conversion is performed on the multi-agent input signal pulse sequence through a sparse matrix generation method based on time coding. Assuming that the multi-agent system includes 8 agents, the time is divided into multiple intervals with 100 milliseconds as a time window. For a pulse of agent A1, if it is generated in the 200-300 millisecond time window, and a related pulse of agent A2 is generated in the 250-350 millisecond time window, the overlapping duration of the two pulse time windows is calculated to be 50 milliseconds, and a matrix with 8 rows and 8 columns is constructed, with the rows and columns corresponding to 8 agents respectively. For the pulses between each two agents, the degree of overlap of their time windows is calculated as the matrix element value. If the overlapping duration of the pulse time windows of agents A1 and A2 is 50 milliseconds and the total time window length is 100 milliseconds, the element value corresponding to row A1 and column A2 in the matrix is ​​50 / 100. =0.5, indicating the temporal correlation between the pulses of the two. By traversing the pulse combinations between all agents in the multi-agent input signal pulse sequence, the matrix elements are calculated and filled, and finally a multi-agent signal sparse matrix with time delay characteristics is generated. This matrix reflects the time delay relationship of each agent pulse in the time dimension.

[0030] Step S14: performing pulse rate sparse matrix conversion on the multi-agent input signal pulse sequence based on the rate coding sparse matrix generation method to generate a multi-agent signal sparse matrix with rate characteristics; In the embodiment of the present invention, a sparse matrix generation method based on rate coding is used to convert the pulse sequence of the multi-agent input signal into a pulse rate sparse matrix to count the number of pulses generated by each agent in a unit time (1 second). As the pulse rate of the agent, agent A1 generates 5 pulses in 1 second, with a pulse rate of 5 times / second; agent A2 generates 3 pulses, with a pulse rate of 3 times / second. A matrix with 8 rows and 8 columns is constructed, with rows and columns corresponding to 8 agents. For the elements in the matrix, the formula is used. (in The first Rank Column element value, For the The pulse rate of each agent, For the If we calculate the matrix element values ​​corresponding to agents A1 and A2, =(5+3) / 4=4, by calculating the pulse rate combinations among all agents and filling the matrix elements, a multi-agent signal sparse matrix with rate characteristics is generated. This matrix reflects the relationship between the pulse rates of each agent and provides a basis for subsequent signal processing.

[0031] Step S15: Based on the multi-agent signal sparse matrix with time delay characteristics and rate characteristics, the multi-agent input signal pulse sequence is subjected to spatiotemporal sparse pulse conversion to generate a multi-agent spatiotemporal sparse pulse sequence.

[0032] In an embodiment of the present invention, a spatiotemporal sparse pulse conversion is performed on the multi-agent input signal pulse sequence based on a previously generated multi-agent signal sparse matrix with time delay characteristics and rate characteristics, and the time screening threshold is set to 0.4 (corresponding to the matrix element value of step S13), and the rate screening threshold is set to 3 (corresponding to the matrix element value of step S14). For a pulse of agent A1, its element value associated with the pulses of other agents in the time delay sparse matrix and the element value corresponding to other agents in the rate sparse matrix are checked. If the corresponding element value of A1 and A3 in the time delay sparse matrix is ​​0.3, which is less than the time delay, the pulse sequence of the multi-agent input signal pulse sequence is converted into a spatiotemporal sparse pulse sequence. The time screening threshold is 0.4; the corresponding element value in the rate sparse matrix is ​​2.5, which is less than the rate screening threshold of 3, then the pulse is judged as a non-critical pulse and is eliminated. By performing the above screening operation on all pulses in the multi-agent input signal pulse sequence, the critical pulses that meet the threshold conditions are retained, and finally a multi-agent spatiotemporal sparse pulse sequence is generated. For example, the original pulse sequence contains 50 pulses. After screening, 20 critical pulses are retained. These pulses have more significant characteristics in the time and rate dimensions, which effectively reduces signal redundancy and improves the privacy protection performance and transmission efficiency of multi-agent communication signals.

[0033] Furthermore, step S13 includes the following steps: Step S131: obtaining corresponding signal pulse time intervals and signal pulse waveform characteristics through a multi-agent input signal pulse sequence; Step S132: Determine the thinning time window based on the signal pulse time interval and the signal pulse waveform characteristics, and obtain the signal pulse thinning time window size related to the time distribution of the original multi-agent signal; Step S133: Based on the signal pulse sparse time window size related to the time distribution of the multi-agent original signal, the pulse amplitude and time series position statistics are performed for each sparse time window moment in the multi-agent input signal pulse sequence to obtain the multi-agent signal pulse amplitude and multi-agent signal time series position corresponding to each sparse time window moment; Step S134: Calculate the delay distribution difference at different moments based on the multi-agent signal pulse amplitude and the multi-agent signal timing position corresponding to each sparse time window moment, and obtain the delay distribution difference between the multi-agent signal pulses at different sparse time windows; Step S135: Based on the delay distribution differences between multi-agent signal pulses at different sparse time windows and combined with the time-coded sparse matrix generation method, the multi-agent input signal pulse sequence is transformed into a delay sparse matrix to generate a multi-agent signal sparse matrix with delay characteristics.

[0034] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 2 Detailed step flow diagram of step S13 in the embodiment, step S13 includes the following steps: Step S131: obtaining corresponding signal pulse time intervals and signal pulse waveform characteristics through a multi-agent input signal pulse sequence; In an embodiment of the present invention, a multi-agent input signal pulse sequence is obtained by taking a network composed of 10 agents as an example. Assume that the signal pulse sequence emitted by agent A1 within 1 second is: pulses are generated at 0.1 seconds, 0.3 seconds, 0.6 seconds, and 0.8 seconds respectively. A high-precision timestamp recording device is used to accurately record the time when each pulse is generated, and the signal pulse time interval is calculated. For example, the time interval between the first pulse and the second pulse is 0.3-0.1=0.2 seconds, and the time interval between the second and third pulses is 0.6-0.3=0.3 seconds. At the same time, a waveform acquisition device such as an oscilloscope is used to obtain the signal pulse waveform characteristics. Taking the voltage signal as an example, the peak voltage, rise time, fall time and other parameters of each pulse are recorded. For example, the first pulse of agent A1 mentioned above has a peak voltage of 5 volts, a rise time of 0.01 seconds, and a fall time of 0.02 seconds. For the input signal pulse sequences of all 10 agents in the network, the corresponding signal pulse time intervals and signal pulse waveform characteristics are obtained in this way to form a complete original signal feature data set.

[0035] Step S132: Determine the thinning time window based on the signal pulse time interval and the signal pulse waveform characteristics, and obtain the signal pulse thinning time window size related to the time distribution of the original multi-agent signal; In the embodiment of the present invention, the dilution time window is determined based on the previously acquired signal pulse time interval and signal pulse waveform characteristics, and the average value of the signal pulse time interval of all agents is calculated using statistical analysis. and standard deviation , assuming that the average time interval of the 10 agent signal pulses is calculated = 0.25 seconds, standard deviation = 0.05 seconds, and the calculation formula for setting the signal pulse sparse time window size is: (in is the adjustment coefficient, which is set according to the signal sparsity requirement, which is also ), then the signal pulse sparse time window size =2×0.25=0.5 seconds. The time window size is related to the time distribution of the original multi-agent signal. In this way, the continuous signal pulse sequence is divided into multiple time windows for subsequent pulse information statistics and analysis, which can not only ensure the capture of key signal features but also reduce the amount of data processing.

[0036] Step S133: Based on the signal pulse sparse time window size related to the time distribution of the multi-agent original signal, the pulse amplitude and time series position statistics are performed for each sparse time window moment in the multi-agent input signal pulse sequence to obtain the multi-agent signal pulse amplitude and multi-agent signal time series position corresponding to each sparse time window moment; In an embodiment of the present invention, the multi-agent input signal pulse sequence is processed based on the previously determined 0.5-second signal pulse sparse time window size. Taking the signal pulse sequence of agent A1 as an example, within the first sparse time window of 0-0.5 seconds, pulses are generated at two times, 0.1 seconds and 0.3 seconds. The amplitudes of these two pulses are counted, assuming they are 5 volts and 4.8 volts respectively, and their timing positions within the time window, that is, the positions of 0.1 seconds and 0.3 seconds relative to time 0, are recorded. For the signal pulse sequences of all 10 agents, the pulse amplitude and timing position statistics are performed at each 0.5-second sparse time window. For example, within the 0.5-1.0 second time window, agent A2 generates a pulse at 0.6 seconds with an amplitude of 5.2 volts. Its amplitude is recorded as 5.2 volts and the timing position is 0.1 second (relative to the 0.5 second moment). Finally, the multi-agent signal pulse amplitude and multi-agent signal timing position corresponding to each sparse time window moment are obtained to form a detailed pulse information statistics table.

[0037] Step S134: Calculate the delay distribution difference at different moments based on the multi-agent signal pulse amplitude and the multi-agent signal timing position corresponding to each sparse time window moment, and obtain the delay distribution difference between the multi-agent signal pulses at different sparse time windows; In this embodiment of the present invention, the delay distribution difference at different moments is calculated based on the multi-agent signal pulse amplitude and multi-agent signal timing position corresponding to each sparse time window moment obtained previously. Taking two adjacent sparse time windows (0-0.5 seconds and 0.6-1.0 seconds) as an example, assuming that agent A1 generates a pulse at 0.1 seconds in the first time window and a pulse at 0.6 seconds in the second time window, the delay between the two pulses is calculated as follows: Seconds, using the variance formula (in is the time delay between different pulses, is the average delay, is the number of pulse logs) to calculate the delay distribution difference. Assume that in two time windows, agents A1, A2, and A3 have two pairs of pulses respectively, and the calculated delays are 0.5 seconds, 0.4 seconds, and 0.6 seconds respectively. The average delay = 0.5+0.4+0.6 / 3=0.5 seconds, so the delay distribution difference = 0+0.01+0.01 / 3≈0.0067. For the pulses of all agents in different sparse time windows, the delay distribution differences are calculated according to this method, and finally comprehensive delay distribution difference data are obtained.

[0038] Step S135: Based on the delay distribution differences between multi-agent signal pulses at different sparse time windows and combined with the time-coded sparse matrix generation method, the multi-agent input signal pulse sequence is transformed into a delay sparse matrix to generate a multi-agent signal sparse matrix with delay characteristics.

[0039] In an embodiment of the present invention, based on the previously obtained delay distribution differences between multi-agent signal pulses at different sparse time window moments, combined with a time-coded sparse matrix generation method, the multi-agent input signal pulse sequence is transformed, and the rows of the sparse matrix are set to represent different agents (a total of 10 rows, corresponding to 10 agents), and the columns represent different sparse time window moments (assuming a total of 20 time window moments, corresponding to 20 columns). For each agent at each time window moment, if there is a pulse, the delay distribution difference value between the pulse and other related pulses is filled in the corresponding position of the matrix; if there is no pulse, 0 is filled in. If agent A1 has a pulse in the first time window and the delay distribution difference with subsequent related pulses is 0.0067, then 0.0067 is filled in the first row and first column of the sparse matrix; if there is no pulse in the second time window, then 0 is filled in the first row and second column. In this way, the entire multi-agent input signal pulse sequence is converted into a 10-row and 20-column multi-agent signal sparse matrix with delay characteristics. This matrix retains the key information of the signal pulse delay and is stored in a sparse form, which facilitates subsequent matrix-based multi-agent privacy protection analysis, such as mining privacy leakage risk points in the signal through matrix operations.

[0040] Furthermore, step S14 includes the following steps: Step S141: performing statistical calculations on the pulse frequency and amplitude of each signal pulse window of the multi-agent input signal pulse sequence to obtain the corresponding multi-agent signal pulse frequency and multi-agent signal pulse amplitude in each signal pulse window; In the embodiment of the present invention, by still taking a network composed of 10 agents as an example, the multi-agent input signal pulse sequence is processed, and the duration of each signal pulse window is set to 0.2 seconds. Taking the signal pulse sequence of agent A1 (generating pulses at 0.1 seconds, 0.3 seconds, 0.6 seconds, and 0.8 seconds respectively) as an example, in the first signal pulse window of 0-0.2 seconds, there is 1 pulse in the window. According to the pulse frequency calculation formula (in is the number of pulses in the window, is the window length), the multi-agent signal pulse frequency in this window can be obtained =5Hz; the pulse amplitude is 5 volts, that is, the multi-agent signal pulse amplitude in this window is 5 volts. For the signal pulse sequence of all 10 agents, the signal pulse window is divided into 0.2 second intervals. For example, agent A2 has 0 pulses in the signal pulse window of 0.2-0.4 seconds, and its pulse frequency is =0Hz, and the pulse amplitude is recorded as 0. In this way, the number and amplitude of pulses of each agent in each signal pulse window are statistically calculated, and finally a complete data set of multi-agent signal pulse frequency and multi-agent signal pulse amplitude corresponding to each signal pulse window is obtained, providing basic data for subsequent analysis.

[0041] Step S142: obtaining a corresponding multi-agent pulse signal strength through the multi-agent input signal pulse sequence, and performing pulse rate modulation on the multi-agent input signal pulse sequence based on the multi-agent pulse signal strength to generate a modulated signal pulse rate corresponding to the multi-agent signal pulse; In the embodiment of the present invention, the multi-agent pulse signal strength is obtained by inputting a multi-agent signal pulse sequence. For agent A1, assuming that its signal pulse amplitudes are 5 volts, 4.8 volts, 5.2 volts, and 5 volts, the formula is used. (in is the signal strength, is the number of pulses, For the The pulse signal strength is calculated by 5+4.8+5.2+5 / 4=5 volts. Similarly, calculate the pulse signal strength of the other 9 agents. Based on the multi-agent pulse signal strength, perform pulse rate modulation on the multi-agent input signal pulse sequence. Set the modulation formula as (in is the modulated pulse rate, is the modulation coefficient, here we set =2), taking agent A1 as an example, its modulated pulse rate =2×5=10 times / second, that is, the modulation signal pulse rate corresponding to the multi-agent signal pulse is generated. The signal pulse sequences of all 10 agents in the network are modulated in this way, and finally the modulation signal pulse rate corresponding to each agent is obtained. The transmission rate of the signal pulse is adjusted to prepare for the construction of a matrix with rate characteristics.

[0042] Step S143: Based on the modulation signal pulse rate corresponding to the multi-agent signal pulse and combined with the rate coding sparse matrix generation method, the multi-agent signal pulse frequency and the multi-agent signal pulse amplitude corresponding to each signal pulse window are sparsely weighted reconstructed to generate a multi-agent signal sparse matrix with rate characteristics.

[0043] In an embodiment of the present invention, based on the modulated signal pulse rate corresponding to the multi-agent signal pulse obtained previously, combined with the sparse matrix generation method of rate coding, the multi-agent signal pulse frequency and the multi-agent signal pulse amplitude corresponding to each signal pulse window obtained are sparsely weighted reconstructed. The rows of the sparse matrix are set to represent different agents (a total of 10 rows, corresponding to 10 agents), and the columns represent different signal pulse window moments (assuming a total of 50 window moments, corresponding to 50 columns). For each agent at each signal pulse window moment, the weighted formula is used. (in The first Rank Column element value, 、 is the weight coefficient, where =0.4, =0.6, For the The agent in the The pulse frequency of the window, For the The agent in the The pulse amplitude of the window, For the The modulation signal pulse rate of each agent, is the maximum value of the modulation signal pulse rate of all agents), assuming that the modulation signal pulse rate of agent A1 =10 times / second, in the first signal pulse window, the pulse frequency =5Hz, pulse amplitude = 5 volts, the maximum modulation signal pulse rate among all agents =12 times / second, then the corresponding element in the matrix =0.4×5+0.6×5×10 / 12=2+2.5=4.5. If an agent has no pulse in a certain window, the corresponding matrix element value is 0. In this way, the pulse frequency, amplitude and modulated pulse rate of each agent in each signal pulse window are weighted and filled into the sparse matrix to generate a 10-row 50-column multi-agent signal sparse matrix with rate characteristics. This matrix integrates the frequency, amplitude and rate information of the multi-agent signal and is stored in a sparse form, which is convenient for subsequent matrix-based multi-agent privacy protection analysis, such as mining potential privacy leakage patterns through matrix feature extraction.

[0044] Furthermore, step S15 includes the following steps: Step S151: performing time-delay sparse filtering on the multi-agent input signal pulse sequence based on the multi-agent signal sparse matrix with time-delay characteristics to obtain a multi-agent time-delay sparse filtered pulse sequence; In this embodiment of the present invention, a time-delay sparse filtering is performed on a multi-agent input signal pulse sequence based on a generated multi-agent signal sparse matrix with time delay characteristics. Assuming the sparse matrix has 10 rows and 20 columns, with rows corresponding to 10 agents and columns corresponding to 20 time windows, the matrix element values ​​represent the differences in the time delay distribution of signal pulses from different agents within the corresponding time window. A filtering threshold of 0.01 is set. For each pulse in the multi-agent input signal pulse sequence, the corresponding element value in the sparse matrix is ​​searched based on the agent and time window it belongs to. For example, the pulse of agent A1 in the third time window is used. If the element value at the corresponding position (row 1, column 3) in the sparse matrix is ​​0, 008, is less than 0.01, then the pulse is retained; if the matrix element value corresponding to a certain agent in a certain time window is greater than or equal to 0.01, such as the element value corresponding to agent A2 in the 5th time window is 0.012, then the pulse is removed. By traversing the pulses of all agents in each time window in the multi-agent input signal pulse sequence, screening according to the above rules, and finally obtaining a multi-agent time-delay sparse filtered pulse sequence, for example, in the original sequence, agents A1-A10 have a total of 100 pulses. After filtering, 70 of them are retained to form a new pulse sequence. This sequence is sparsely processed in the time dimension, reducing redundant pulses and highlighting key pulses with obvious delay characteristics.

[0045] Step S152: performing rate sparse modulation on the multi-agent time-delay sparse filtered pulse sequence in spatial position distribution based on the multi-agent signal sparse matrix with rate characteristics to generate a multi-agent rate sparse modulated pulse sequence; In an embodiment of the present invention, a multi-agent time-delay sparse filter pulse sequence obtained previously is subjected to rate sparse modulation in terms of spatial position distribution based on a multi-agent signal sparse matrix with rate characteristics. The rate characteristic sparse matrix is ​​10 rows and 50 columns, where a row represents 10 agents and a column represents 50 signal pulse window moments. The matrix element values ​​are calculated by weighting the pulse frequency, amplitude and modulated pulse rate. For each pulse in the multi-agent time-delay sparse filter pulse sequence, the element value at the corresponding position in the sparse matrix is ​​obtained according to the agent to which it belongs and the window moment. Taking the pulse of agent A3 at the 10th window moment as an example, assuming that the pulse is in the rate characteristic sparse matrix The corresponding element value in the array is 3.8, and the modulation rule is set: when the element value is greater than 3, the spatial transmission path of the pulse is adjusted to a path with a higher priority; when the element value is less than or equal to 3, the original transmission path is maintained, so the transmission path of the pulse is adjusted to a line with a higher priority for transmission. For all pulses in the multi-agent time-delay sparse filtering pulse sequence, this rule is followed, and the spatial transmission path is adjusted according to the element value of the rate characteristic sparse matrix, and finally a multi-agent rate sparse modulated pulse sequence is generated. In this way, the pulse is modulated according to the rate characteristics of the signal in the spatial position distribution, the transmission path of the pulse is optimized, and the efficiency and pertinence of signal transmission are improved.

[0046] Step S153: Obtain the pulse energy corresponding to each pulse signal through the multi-agent rate sparse modulation pulse sequence, and perform pulse attenuation evaluation based on the pulse energy corresponding to each pulse signal to obtain the pulse attenuation factor corresponding to the multi-agent pulse signal; In the embodiment of the present invention, the pulse energy corresponding to each pulse signal is obtained by multi-agent rate sparse modulation pulse sequence. For a pulse of agent A1, assuming that its amplitude is 4.5 volts and its duration is 0.01 second, according to the pulse energy calculation formula (Assuming that the pulse signal can be equivalent to a capacitor discharge model, the capacitor Farad), the pulse energy can be obtained Joule, for each pulse of all agents in the multi-agent rate sparse modulation pulse sequence, the energy size is calculated according to this formula, and the pulse attenuation evaluation is performed according to the pulse energy size corresponding to each pulse signal. The attenuation evaluation formula is set as (in is the pulse decay factor, is the minimum value among all pulse energies), assuming that the minimum value among all pulse energies Joule, for the pulse of the above agent A1, its pulse attenuation factor is By calculating the attenuation factor for all pulses, we can finally obtain the set of pulse attenuation factors corresponding to the multi-agent pulse signal. This factor reflects the attenuation degree of each pulse relative to the minimum energy pulse.

[0047] Step S154: Based on the pulse attenuation factor corresponding to the multi-agent pulse signal, the spatiotemporal pulse distribution of each corresponding pulse signal in the multi-agent rate sparse modulation pulse sequence is optimized to generate a multi-agent spatiotemporal sparse pulse sequence.

[0048] In an embodiment of the present invention, the spatiotemporal pulse distribution of each pulse signal in the multi-agent rate sparse modulation pulse sequence is optimized based on the pulse attenuation factor corresponding to the multi-agent pulse signal, and an optimization rule is set: when the pulse attenuation factor is less than 0.8, the pulse is delayed backward in time by 0.02 seconds and spatially adjusted to the backup transmission path; when the pulse attenuation factor is greater than or equal to 0.8, the spatiotemporal position of the pulse is kept unchanged. Taking a pulse of agent A4 as an example, its pulse attenuation factor is 0.75. According to the rule, the pulse is delayed from the original 0.5 second to 0.52 on the time axis. At the same time, at the same time, the transmission path is spatially adjusted from the original transmission path to the backup path for transmission; for the pulse with an attenuation factor of 0.85 for agent A5, its spatiotemporal position is maintained. All pulses in the multi-agent rate sparse modulation pulse sequence are adjusted in time and space according to their respective pulse attenuation factors according to this rule, and finally a multi-agent spatiotemporal sparse pulse sequence is generated. This sequence optimizes the distribution of pulses in time and space dimensions, reduces the interference of pulses with weaker energy, enhances the stability of signal transmission and the privacy protection performance, and makes the multi-agent communication signal more reasonable and efficient at the spatiotemporal level.

[0049] Furthermore, step S2 includes the following steps: Step S21: Constructing a corresponding hierarchical spatiotemporal spike neural network by adopting a structure combining a recursive neural network and a long short-term memory network, specifically including a bottom layer, a middle layer, and a top layer, wherein the bottom layer is used to process the original pulse sequence, the middle layer is used to extract the corresponding pulse timing features, and the top layer is used to decide and output the corresponding communication pulse sequence; In an embodiment of the present invention, a hierarchical spatiotemporal spike neural network is constructed, and a structure combining a recursive neural network and a long short-term memory network is adopted. The bottom layer consists of 10 neurons for processing the original pulse sequence. Each neuron corresponds to an intelligent agent and receives the multi-agent spatiotemporal sparse pulse sequence as input. The middle layer contains 20 neurons for extracting pulse timing features. Each neuron is connected to all neurons in the bottom layer. The top layer is provided with 5 neurons for making decisions on output communication pulse sequences, receiving the output of the middle layer and generating the final result. The bottom layer neurons adopt the IF (Integrate-and-Fire) model, and its membrane potential update formula is: ,in =0.8 is the membrane potential time constant, is the connection weight, is the input pulse, is the threshold. When the middle-layer neurons receive pulses from the bottom layer, the timing information is processed by the LSTM unit. The calculation formulas of its input gate, forget gate and output gate are: 、 、 ,in is the sigmoid function, is the weight matrix, is a bias term. The top-level neuron uses the softmax function to make decisions based on the temporal features extracted by the middle layer. The formula is: ,in For the The input of a neuron, is the number of output categories.

[0050] Step S22: Input the multi-agent spatiotemporal sparse pulse sequence into the bottom layer of the hierarchical spatiotemporal spiking neural network for pulse preprocessing, and input the preprocessed multi-agent spatiotemporal sparse pulse sequence into the middle layer of the hierarchical spatiotemporal spiking neural network to train the pulse extraction network by introducing a pulse timing attention mechanism. At the same time, a differentially private spike network training method is designed and constructed during the training process to introduce Gaussian noise into the spike neural network and combine the gradient clipping strategy of the training process to prevent gradient reversal attacks, so as to train and generate the pulse timing feature output corresponding to each agent; In this embodiment of the present invention, a multi-agent spatiotemporal sparse pulse sequence is input into the bottom layer of a hierarchical spatiotemporal pulse neural network for preprocessing. The bottom layer neurons integrate the input pulses and generate output pulses when the membrane potential exceeds the threshold. The preprocessed pulse sequence is passed to the middle layer, and the middle layer introduces a pulse timing attention mechanism to train the pulse extraction network. The attention weight calculation formula is: ,in , is the hidden state at the current moment, For the The characteristic vector of the pulse, 、 、 For a trainable weight matrix, during the training process, a pulse network training method based on differential privacy is designed. At each parameter update, Gaussian noise is added to the gradient, and the standard deviation of the noise is ,in is the sensitivity of the gradient, = 0.5 is the privacy budget parameter, and the gradient clipping strategy is adopted to clip the gradient norm to =1.0, to prevent gradient reversal attacks, the training objective function is the cross entropy loss function: ,in is the true label, To predict the probability, the network parameters are updated by the stochastic gradient descent algorithm, and the learning rate is set to = 0.01, and after 100 training cycles, the pulse timing feature output corresponding to each agent is finally generated.

[0051] Step S23: The pulse interaction frequency and pulse sensitivity between each agent are calculated by outputting the pulse timing characteristics corresponding to each agent at the top layer of the hierarchical spatiotemporal pulse neural network, and the corresponding privacy budget is dynamically configured according to the pulse interaction frequency and pulse sensitivity between each agent. parameters, and according to the privacy budget The parameters in the top layer decide the corresponding communication pulse sequence output to generate the multi-agent timing-critical communication pulse sequence.

[0052] In this embodiment of the present invention, the pulse interaction frequency and pulse sensitivity between agents are calculated based on the pulse timing characteristics output by each agent at the top level of the hierarchical spatiotemporal pulse neural network. The pulse interaction frequency calculation formula is: ,in and Agents and At the moment The pulse output, is the total number of time steps, and the pulse sensitivity calculation formula is: ,in The privacy budget parameter is dynamically configured according to the pulse interaction frequency and pulse sensitivity as the loss function. For pairs of agents with high interaction frequency and high sensitivity, a smaller privacy budget parameter is assigned. The calculation formula is: ,in =1.0 is the initial privacy budget. When the top-level decision outputs the communication pulse sequence, noise is injected into the output according to the privacy budget parameter to generate the multi-agent timing-critical communication pulse sequence. The noise injection formula is: ,in is the original output, , In this way, the privacy information between intelligent agents is effectively protected while ensuring the communication quality.

[0053] Furthermore, step S3 includes the following steps: Step S31: performing pulse timing disturbance analysis on the multi-agent timing-critical communication pulse sequence to obtain the multi-agent communication pulse timing disturbance distribution characteristics; In an embodiment of the present invention, a pulse timing disturbance analysis is performed on a multi-agent timing-critical communication pulse sequence. Assuming that there are five agents (A, B, C, D, and E) in a multi-agent system, a pulse sequence of agent A is taken as an example. The pulse generation times are 0.1 seconds, 0.3 seconds, 0.5 seconds, 0.7 seconds, and 0.9 seconds, respectively. By calculating the time intervals between adjacent pulses, the time interval sequence is 0.2 seconds, 0.2 seconds, 0.2 seconds, and 0.2 seconds. The statistical analysis method is used to calculate the average value of these time intervals. and standard deviation ,average value =0.2+0.2+0.2+0.2 / 4=0.2 seconds, standard deviation = 0 seconds, which shows that the pulse sequence of agent A is relatively stable and the disturbance is small. The same analysis is performed on the pulse sequences of the other four agents. The pulse time interval sequence of agent B is 0.15 seconds, 0.25 seconds, 0.18 seconds, and 0.22 seconds. The average value is calculated. = 0.2 seconds, standard deviation ≈0.035 seconds, indicating that there is a certain disturbance in the pulse timing of agent B. By analyzing the pulse sequences of all agents, the distribution characteristics of the pulse timing disturbance of multi-agent communication are obtained, that is, the degree of disturbance of the pulse timing of different agents is different. The pulse timing of some agents is relatively stable, while the pulse timing of some agents has large fluctuations.

[0054] Step S32: Obtain the pulse communication disturbance delay and pulse communication disturbance frequency difference between each agent through the multi-agent communication pulse timing disturbance distribution characteristics, and perform pulse time jitter amplitude statistics on each pulse signal in the multi-agent timing critical communication pulse sequence based on the pulse communication disturbance delay and pulse communication disturbance frequency difference between each agent to obtain the communication pulse signal time jitter amplitude between each agent; In an embodiment of the present invention, the pulse communication disturbance delay and the pulse communication disturbance frequency difference between each agent are obtained based on the previously obtained multi-agent communication pulse timing disturbance distribution characteristics. Taking agents A and B as an example, assuming that a pulse of agent A is generated at 0.3 seconds and a related pulse of agent B is generated at 0.35 seconds, the pulse communication disturbance delay is 0.35-0.3=0.05 seconds, and the pulse communication disturbance frequency difference is calculated. Taking 1 second as the time window, the number of pulses of agents A and B in the time window is counted. Assuming that agent A generates 5 pulses in 1 second and agent B generates 4 pulses in 1 second, the pulse communication disturbance frequency difference is |5-4|=1 time / second. Based on the pulse communication disturbance delay and the pulse communication disturbance frequency difference between each agent, the pulse time jitter amplitude statistics are performed on each pulse signal in the multi-agent timing critical communication pulse sequence, and the jitter amplitude calculation formula is set as ,in is the pulse communication disturbance delay, is the frequency difference of pulse communication disturbance. Taking a pulse pair between agents A and B as an example, = 0.05 seconds, =1 time / second, then the pulse time jitter amplitude =1.05, and by calculating the pulse pairs between all agents, the time jitter amplitude of the communication pulse signal between each agent is obtained. For example, the pulse time jitter amplitude between agents A and C is 0.8, and the pulse time jitter amplitude between agents B and D is 1.2, etc.

[0055] Step S33: Generate a corresponding multi-agent pulse time jitter amplitude spectrum based on the time jitter amplitude of the communication pulse signals between each agent, and perform pulse selective deletion optimization on each pulse signal in the multi-agent timing-critical communication pulse sequence based on the multi-agent pulse time jitter amplitude spectrum to filter out redundant pulse signals corresponding to the consistent time jitter amplitude of the communication pulse signals, thereby generating a multi-agent pulse optimized deletion sequence; In an embodiment of the present invention, a corresponding multi-agent pulse time jitter amplitude spectrum is generated based on the previously obtained communication pulse signal time jitter amplitude between each agent, and a bar graph is drawn with the agent as the horizontal axis and the pulse time jitter amplitude as the vertical axis. For example, the pulse time jitter amplitude of agent A is 0.5, the pulse time jitter amplitude of agent B is 0.8, the pulse time jitter amplitude of agent C is 0.6, the pulse time jitter amplitude of agent D is 0.9, and the pulse time jitter amplitude of agent E is 0.7. On the spectrum, pulse signals with consistent pulse time jitter amplitudes are marked. Assuming that the pulse time jitter amplitudes of agents A and C are both 0.6 in a certain time period, these pulse signals are redundant pulse signals. Based on the multi-agent pulse time jitter amplitude spectrum, pulse selective deletion optimization is performed on each pulse signal in the multi-agent timing-critical communication pulse sequence, and redundant pulse signals corresponding to the consistent communication pulse signal time jitter amplitude are screened and deleted. By traversing the pulse sequence, the time jitter amplitude of each pulse is compared with the data in the spectrum, and the redundant pulses are deleted. For example, in a sequence containing 100 pulses, 20 redundant pulses were deleted after comparison and screening, and finally a multi-agent pulse optimized deletion sequence was generated, which reduced redundant information and improved the effectiveness and privacy protection performance of the pulse sequence.

[0056] Step S34: Perform pulse perturbation distribution randomization processing on the timing of the corresponding remaining pulse signals in the multi-agent pulse optimization deletion sequence to generate a multi-agent pulse perturbation randomization sequence.

[0057] In an embodiment of the present invention, the pulse perturbation distribution randomization processing is performed on the timing of the corresponding remaining pulse signals in the previously generated multi-agent pulse optimization deletion sequence, and the randomization range is set to ±0.1 seconds. Taking a remaining pulse of agent A as an example, its original generation time is 0.5 seconds. A random number is generated within the randomization range. Assuming it is 0.05 seconds, the new generation time of the pulse is 0.5+0.05=0.55 seconds. All the remaining pulse signals in the multi-agent pulse optimization deletion sequence are randomized in this way. For example, a The original pulse time is 0.7 seconds, and the random number is -0.08 seconds, so the new time is 0.7-0.08=0.62 seconds. By randomizing the timing of all remaining pulses, a multi-agent pulse perturbation randomized sequence is generated. The temporal distribution of this sequence is more random, which increases the unpredictability of the pulse sequence and further improves the privacy protection performance of multi-agent communication. For example, the time distribution of pulses in the original optimized deletion sequence is relatively concentrated. After randomization, the temporal distribution of pulses is more dispersed, reducing the risk of obtaining privacy information by analyzing the pulse timing.

[0058] Furthermore, step S4 includes the following steps: Step S41: constructing privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, membership inference attack, and attribute inference attack; In an embodiment of the present invention, when constructing a pulse sequence reconstruction attack scenario model, it is assumed that the attacker can obtain a randomized sequence of multi-agent pulse perturbations, and through known pulse generation rules and communication protocols, use a reconstruction algorithm based on maximum likelihood estimation to try to restore the original pulse sequence. Taking a system containing four agents (A, B, C, D) as an example, assuming that the original pulse sequence is binary coded, the attacker calculates the probability of generating the current perturbation sequence under different original sequence assumptions based on the received perturbation sequence, and selects the original sequence with the highest probability as the reconstruction result. In the member reasoning attack scenario model, the attacker uses a similarity calculation method to determine whether a certain agent belongs to the system based on the historical pulse sequence data of the multi-agent system and the currently observed randomized pulse perturbation sequence. For example, using the cosine similarity formula , calculate the target agent pulse sequence Average pulse sequence of agents in the system The similarity threshold is set to 0.7. If the threshold is exceeded, it is inferred that the target agent belongs to the system. In the attribute reasoning attack scenario model, the attacker builds a classification model based on a decision tree to infer the attributes of the agent based on the statistical characteristics of the agent's pulse sequence (such as pulse frequency, pulse interval distribution, etc.). Assuming that the agent has two attributes, "working mode 1" and "working mode 2", the attacker trains a decision tree model by analyzing a large amount of agent pulse sequence data with known attributes. When a new pulse perturbation randomized sequence is obtained, the model is input to predict the agent's attributes.

[0059] Step S42: Monte Carlo simulation is used in combination with the corresponding privacy leakage attack scenario model to simulate the privacy leakage attack on the randomized sequence of multi-agent pulse perturbations to generate the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios; In an embodiment of the present invention, a privacy leakage attack simulation is performed on a multi-agent pulse perturbation randomized sequence by using Monte Carlo simulation combined with a constructed privacy leakage attack scenario model. The number of Monte Carlo simulations is set to 10,000 times, and each attack scenario model is simulated independently. In the pulse sequence reconstruction attack simulation, in each simulation, the attacker reconstructs the original pulse sequence based on the current multi-agent pulse perturbation randomized sequence using the maximum likelihood estimation reconstruction algorithm. Taking the pulse sequence of agent A as an example, if the original sequence is [1, 0, 1, 0, 1], the simulated reconstruction results in [1, 1, 1, 0, 0]. The original sequence and the reconstructed sequence are compared, and the number of reconstructed erroneous pulses is counted. After 10,000 simulations, the ratio of the number of reconstructed erroneous pulses to the total number of pulses is calculated as the privacy leakage probability under this attack scenario. Assuming that the calculation result It is 0.15. For the member reasoning attack simulation, an agent is randomly selected as the target in each simulation, and the cosine similarity algorithm is used to determine whether it belongs to the system. If it actually belongs to the system but is mistakenly judged as not belonging, or if it actually does not belong but is mistakenly judged as belonging, it is recorded as a privacy leak. After 10,000 simulations, the number of privacy leaks is counted and divided by the total number of simulations to obtain the privacy leakage probability under the member reasoning attack scenario, which is assumed to be 0.1. In the attribute reasoning attack simulation, each simulation inputs the pulse perturbation randomization sequence into the trained decision tree classification model for attribute prediction. If the prediction is wrong, it is regarded as a privacy leak. After 10,000 simulations, the proportion of the number of prediction errors to the total number of simulations is calculated. It is assumed that the privacy leakage probability under the attribute reasoning attack scenario is 0.12, thereby generating the corresponding multi-agent pulse privacy leakage probability under different attack scenarios.

[0060] Step S43: Based on the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios, the privacy leakage diffusion impact analysis is performed on the multi-agent pulse perturbation randomized sequence to obtain the corresponding diffusion path and impact range of the corresponding pulse privacy leakage event in the multi-agent system; In an embodiment of the present invention, based on the probability of privacy leakage of multi-agent pulses under different attack scenarios, the privacy leakage diffusion impact of the randomized sequence of multi-agent pulse disturbances is analyzed. Taking the privacy leakage caused by the pulse sequence reconstruction attack as an example, it is assumed that the original pulse sequence of agent A contains key task instruction information. When the sequence is reconstructed incorrectly, due to the data interaction dependency relationship between agent A and agents B and C, the erroneous instruction information will be transmitted to agents B and C, resulting in deviations in their task execution, which in turn affects agent D that has data interaction with B and C, forming a diffusion path of A→B→C→D. By analyzing the data interaction topology and dependency relationship between the agents in the system, the privacy leakage caused by the pulse sequence reconstruction attack is analyzed. The influence degree of each agent is determined by combining the privacy leakage probability and setting the influence degree calculation formula as I=p×d, where p is the privacy leakage probability under the corresponding attack scenario and d is the dependence degree of the agent in the data interaction topology structure (obtained by calculating the sum of the weights of the connection edges between the agent and other affected agents). Assuming that the dependence degree of agent B is d=3 and p=0.15 under the pulse sequence reconstruction attack scenario, its influence degree is I=0.15×3=0.45. The influence range threshold is set to 0.3, and the agents with an influence degree greater than or equal to the threshold are included in the influence range. Finally, the diffusion path and influence range of the corresponding pulse privacy leakage event in the multi-agent system are obtained.

[0061] Step S44: Constructing a privacy-efficiency optimization target corresponding to the privacy protection level, communication bandwidth consumption, and computational delay based on the diffusion path and impact range of the corresponding pulse privacy leakage event in the multi-agent system; In this embodiment of the present invention, a privacy-efficiency optimization goal is constructed based on the diffusion path and impact range of the pulse privacy leakage event. The privacy protection level is measured by the inverse of the weighted sum of the privacy leakage probability under different attack scenarios. The weights of the pulse sequence reconstruction attack, membership inference attack, and attribute inference attack are respectively =0.4, =0.3, =0.3, the corresponding privacy leakage probability is 、 、 , then the privacy protection level , the communication bandwidth consumption is determined by calculating the bandwidth resources occupied by the randomized sequence of multi-agent pulse perturbations during the transmission process, assuming that each pulse signal occupies bit bandwidth, the number of pulses in the sequence is , the transmission time is , then the communication bandwidth consumption , the computational delay is measured by the time required for the multi-agent system to process a pulse signal. Suppose the average time it takes for the system to process a single pulse signal is , then calculate the delay = , thereby constructing a privacy-efficiency optimization target corresponding to the privacy protection level, communication bandwidth consumption and computing delay, aiming to balance these three indicators and improve the performance of the multi-agent system.

[0062] Step S45: Based on the privacy-efficiency optimization objectives corresponding to the privacy protection level, communication bandwidth consumption and computing delay, a privacy leakage protection analysis is performed on the randomized sequence of multi-agent pulse perturbations, and the Pareto optimal solution set corresponding to the privacy-efficiency optimization objectives is solved by the NSGA-II algorithm. Based on the Pareto optimal solution set, a balance strategy between privacy and efficiency corresponding to the pulse signal communication is obtained to generate a privacy-efficiency protection optimization strategy corresponding to the multi-agent pulse communication.

[0063] In an embodiment of the present invention, based on the privacy-efficiency optimization goal, the NSGA-II algorithm is used to solve the corresponding Pareto optimal solution set. The parameters of the multi-agent pulse perturbation randomization sequence (such as the threshold in the pulse generation rule, the range of randomization processing, etc.) are used as decision variables and input into the NSGA-II algorithm. The algorithm generates multiple groups of pulse perturbation randomization sequence schemes with different parameter combinations through genetic operations such as selection, crossover, and mutation. For each group of schemes, its corresponding privacy protection level, communication bandwidth consumption, and computing delay are calculated. Non-dominated sorting and congestion calculation are used to screen out non-dominated solutions, and the Pareto optimal solution set is gradually constructed. Assuming that After multiple rounds of iterative calculations, a Pareto optimal solution set containing 10 groups of solutions was obtained. A suitable solution was selected from the Pareto optimal solution set as the balance strategy between privacy and efficiency corresponding to pulse signal communication. For example, the privacy protection level of Solution 1 is -0.12, the communication bandwidth consumption is 10Mbps, and the calculation delay is 50ms; the privacy protection level of Solution 2 is -0.1, the communication bandwidth consumption is 12Mbps, and the calculation delay is 45ms. According to actual application requirements, if privacy protection is more important, Solution 1 is selected; if communication efficiency is higher, Solution 2 is selected. Finally, a privacy-efficiency protection optimization strategy corresponding to multi-agent pulse communication is generated.

[0064] Furthermore, step S42 includes the following steps: Monte Carlo simulation is used to combine the privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, membership inference attack and attribute inference attack to simulate and infer the pulse attack of the randomized multi-agent pulse perturbation sequence, so as to generate the corresponding multi-agent pulse perturbation attack generation process under different attack scenarios. In an embodiment of the present invention, a Monte Carlo simulation is used in combination with privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, membership inference attack and attribute inference attack to perform pulse attack simulation inference on a multi-agent pulse perturbation randomized sequence. Assuming that there are 6 agents (A, B, C, D, E, F) in the multi-agent system, the number of Monte Carlo simulations is set to 10,000 times. For the pulse sequence reconstruction attack scenario model, in each simulation, the attacker attempts to reconstruct the original pulse sequence based on the received multi-agent pulse perturbation randomized sequence through a specific algorithm (such as an algorithm based on Bayesian inference). Taking agent A as an example, assuming that its original pulse sequence is [0, 1, 0, 1, 0]. In the simulation, the attacker uses the algorithm to reconstruct the perturbation randomized sequence received. After multiple simulations, the degree of difference between the result of each reconstruction and the original sequence is recorded to evaluate the effectiveness of the pulse sequence reconstruction attack. For the membership inference attack scenario model, the attacker attempts to infer whether a certain agent belongs to the multi-agent system by analyzing the pulse sequence. In each simulation, the attacker replaces an unknown agent with the original sequence. The pulse sequence is compared with a known randomized multi-agent pulse perturbation sequence (for example, using the cosine similarity algorithm). Assume that in one simulation, the pulse sequence of the unknown agent has a similarity of 0.8 with that of agent B. Through multiple simulations, the number of times the unknown agent is judged to belong to the system is statistically analyzed to evaluate the success rate of the membership inference attack. In the attribute inference attack scenario model, the attacker attempts to infer certain attributes of the agent (such as the agent's working status) from the pulse sequence. In each simulation, the attacker uses a classification algorithm (such as a decision tree classification algorithm) based on the characteristics of the pulse sequence (such as pulse frequency and pulse amplitude) to infer the agent's attributes. Assume that in one simulation, based on the pulse sequence characteristics of agent C, its working status is inferred to be "busy." Through multiple simulations, the number of correct inferences is statistically analyzed to evaluate the accuracy of the attribute inference attack. Through 10,000 Monte Carlo simulations, the corresponding multi-agent pulse perturbation attack generation process for different attack scenarios is finally generated. The attack attempts, results, and related parameters in each simulation are recorded in detail, providing a rich data foundation for subsequent analysis.

[0065] Preferably, multiple rounds of reverse reasoning of pulse decision are performed on the multi-agent pulse perturbation attack generation process corresponding to different attack scenarios to generate a multi-agent pulse perturbation attack decision tree corresponding to different attack scenarios; In an embodiment of the present invention, by performing multiple rounds of reverse reasoning on the pulse decision-making of the multi-agent pulse disturbance attack generation process corresponding to different attack scenarios, taking the pulse sequence reconstruction attack scenario as an example, reverse reasoning is started from the reconstruction result of each simulation. It is assumed that in one simulation, the reconstruction result is different from the original sequence. The reasons leading to the difference are analyzed, such as the loss or incorrect reconstruction of a certain pulse. These reasons are used as nodes of the decision tree. For example, node A represents "pulse loss" and node B represents "pulse error reconstruction". Then, under what circumstances these reasons will occur, it is further analyzed. For example, node A can be divided into child nodes A1 and A2, which represent "noise loss". "Acoustic interference causes pulse loss", "A2" means "algorithm defects cause pulse loss", etc. Through reverse reasoning of multiple simulation results, a complete multi-agent pulse perturbation attack decision tree is gradually constructed. For member reasoning attack scenarios and attribute reasoning attack scenarios, the same method is used for reverse reasoning and decision tree construction. For example, in the member reasoning attack decision tree, node C means "similarity is higher than the threshold", sub-node C1 means "belongs to the system", and C2 means "does not belong to the system", etc. These decision trees clearly show the logic and possible paths of attack decisions in different attack scenarios, providing an intuitive model for subsequent analysis.

[0066] Preferably, the multi-agent pulse perturbation stagnation time points corresponding to different attack scenarios are obtained through the multi-agent pulse perturbation attack decision trees corresponding to different attack scenarios, and the leakage probability prediction calculation is performed on the multi-agent pulse perturbation attack decision trees corresponding to different attack scenarios based on the multi-agent pulse perturbation stagnation time points corresponding to different attack scenarios to generate the multi-agent pulse privacy leakage probability corresponding to different attack scenarios.

[0067] In an embodiment of the present invention, the multi-agent pulse perturbation stagnation time points corresponding to different attack scenarios are obtained by using the multi-agent pulse perturbation attack decision tree corresponding to different attack scenarios. Taking the pulse sequence reconstruction attack decision tree as an example, the path with the smallest difference between the reconstruction result and the original sequence is found in the tree. The time corresponding to a certain node on the path is the pulse perturbation stagnation time point. Assuming that on a certain path, the time corresponding to node D is the 5th simulation, and the difference between the reconstruction result and the original sequence is the smallest at this time, then the time point corresponding to the 5th simulation is the pulse perturbation stagnation time point. Based on the multi-agent pulse perturbation stagnation time points corresponding to different attack scenarios, the leakage probability prediction calculation is performed on the multi-agent pulse perturbation attack decision tree corresponding to different attack scenarios, and the leakage probability calculation formula is set to P = n / N, where n is the number of successful privacy leaks after the pulse perturbation stops, and N is the total number of simulations. Taking the pulse sequence reconstruction attack scenario as an example, assuming that the original pulse sequence is successfully reconstructed 100 times after the pulse perturbation stops, and the total number of simulations is 10,000 times, the leakage probability P=100 / 10,000=0.01. For the membership inference attack scenario and the attribute inference attack scenario, the same formula is used for calculation. For example, in the membership inference attack scenario, after the pulse perturbation stops, the number of successful inferences of the agent membership is 200 times, then the leakage probability P=200 / 10,000=0.02; in the attribute inference attack scenario, the number of successful inferences of the agent attributes is 150 times, then the leakage probability P=150 / 10,000=0.015, and finally the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios are obtained.

[0068] Furthermore, the present invention also provides a multi-agent privacy protection analysis system for an impulse communication mechanism, which is used to perform the multi-agent privacy protection analysis method for the impulse communication mechanism described above. The multi-agent privacy protection analysis system for the impulse communication mechanism includes: A spatiotemporal sparse pulse conversion module is used to obtain a continuous input signal sequence corresponding to the multi-agent system and perform spatiotemporal sparse pulse conversion on the continuous input signal sequence corresponding to the multi-agent system based on a sparse matrix generation method of time coding and rate coding to generate a multi-agent spatiotemporal sparse pulse sequence; The pulse communication decision output module is used to construct a hierarchical spatiotemporal pulse neural network and input the multi-agent spatiotemporal sparse pulse sequence into the hierarchical spatiotemporal pulse neural network. Combined with the pulse network training method based on differential privacy, the pulse communication decision output is performed to generate the multi-agent time-critical communication pulse sequence; The pulse perturbation randomization module is used to perform pulse perturbation randomization processing on the multi-agent timing-critical communication pulse sequence based on the combination of pulse time jitter and pulse deletion, thereby generating a multi-agent pulse perturbation randomized sequence; The pulse privacy protection analysis module is used to construct privacy leakage attack scenario models corresponding to pulse sequence reconstruction attacks, member inference attacks, and attribute inference attacks, and uses Monte Carlo simulation combined with the corresponding privacy leakage attack scenario models to simulate privacy leakage attacks on multi-agent pulse perturbation randomized sequences to generate the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios; based on the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios, the privacy leakage protection analysis of the multi-agent pulse perturbation randomized sequences is performed to generate the privacy-efficiency protection optimization strategy corresponding to multi-agent pulse communication.

[0069] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0070] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A multi-agent privacy protection analysis method for pulse communication mechanism, characterized by: The following steps are involved: Step S1: obtaining a continuous input signal sequence corresponding to the multi-agent system, and performing spatiotemporal sparse pulse conversion on the continuous input signal sequence corresponding to the multi-agent system based on a sparse matrix generation method of time coding and rate coding to generate a multi-agent spatiotemporal sparse pulse sequence; Step S2: Construct a hierarchical spatiotemporal pulse neural network, and input the multi-agent spatiotemporal sparse pulse sequence into the hierarchical spatiotemporal pulse neural network. Combined with the pulse network training method based on differential privacy, the pulse communication decision output is performed to generate the multi-agent time-critical communication pulse sequence. Step S3: performing pulse perturbation randomization processing based on a combination of pulse time jitter and pulse deletion on the multi-agent timing-critical communication pulse sequence to generate a multi-agent pulse perturbation randomized sequence; Step S4: Construct privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, membership inference attack, and attribute inference attack, and use Monte Carlo simulation combined with the corresponding privacy leakage attack scenario model to simulate the privacy leakage attack on the multi-agent pulse perturbation randomized sequence to generate the corresponding multi-agent pulse privacy leakage probability under different attack scenarios; Based on the corresponding multi-agent pulse privacy leakage probability under different attack scenarios, the privacy leakage protection analysis of the multi-agent pulse perturbation randomized sequence is performed to generate the privacy-efficiency protection optimization strategy corresponding to the multi-agent pulse communication.

2. The multi-agent privacy protection analysis method for the pulse communication mechanism according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: Obtain a continuous input signal sequence corresponding to the multi-agent system; Step S12: performing input signal pulse generation on the continuous input signal sequence corresponding to the multi-agent system to generate a multi-agent input signal pulse sequence; Step S13: performing a time-delay sparse matrix conversion on the multi-agent input signal pulse sequence based on a sparse matrix generation method of time coding to generate a multi-agent signal sparse matrix with time delay characteristics; Step S14: performing pulse rate sparse matrix conversion on the multi-agent input signal pulse sequence based on the rate coding sparse matrix generation method to generate a multi-agent signal sparse matrix with rate characteristics; Step S15: Based on the multi-agent signal sparse matrix with time delay characteristics and rate characteristics, the multi-agent input signal pulse sequence is subjected to spatiotemporal sparse pulse conversion to generate a multi-agent spatiotemporal sparse pulse sequence.

3. The multi-agent privacy protection analysis method for the pulse communication mechanism according to claim 2 is characterized in that: Step S13 includes the following steps: Step S131: obtaining corresponding signal pulse time intervals and signal pulse waveform characteristics through a multi-agent input signal pulse sequence; Step S132: Determine the thinning time window based on the signal pulse time interval and the signal pulse waveform characteristics, and obtain the signal pulse thinning time window size related to the time distribution of the original multi-agent signal; Step S133: Based on the signal pulse sparse time window size related to the time distribution of the multi-agent original signal, the pulse amplitude and time series position statistics are performed for each sparse time window moment in the multi-agent input signal pulse sequence to obtain the multi-agent signal pulse amplitude and multi-agent signal time series position corresponding to each sparse time window moment; Step S134: Calculate the delay distribution difference at different moments based on the multi-agent signal pulse amplitude and the multi-agent signal timing position corresponding to each sparse time window moment, and obtain the delay distribution difference between the multi-agent signal pulses at different sparse time windows; Step S135: Based on the delay distribution differences between multi-agent signal pulses at different sparse time windows and combined with the time-coded sparse matrix generation method, the multi-agent input signal pulse sequence is transformed into a delay sparse matrix to generate a multi-agent signal sparse matrix with delay characteristics.

4. The multi-agent privacy protection analysis method for the pulse communication mechanism according to claim 2 is characterized in that: Step S14 includes the following steps: Step S141: performing statistical calculations on the pulse frequency and amplitude of each signal pulse window of the multi-agent input signal pulse sequence to obtain the corresponding multi-agent signal pulse frequency and multi-agent signal pulse amplitude in each signal pulse window; Step S142: obtaining a corresponding multi-agent pulse signal strength through the multi-agent input signal pulse sequence, and performing pulse rate modulation on the multi-agent input signal pulse sequence based on the multi-agent pulse signal strength to generate a modulated signal pulse rate corresponding to the multi-agent signal pulse; Step S143: Based on the modulation signal pulse rate corresponding to the multi-agent signal pulse and combined with the rate coding sparse matrix generation method, the multi-agent signal pulse frequency and the multi-agent signal pulse amplitude corresponding to each signal pulse window are sparsely weighted reconstructed to generate a multi-agent signal sparse matrix with rate characteristics.

5. The multi-agent privacy protection analysis method for the pulse communication mechanism according to claim 2 is characterized in that: Step S15 includes the following steps: Step S151: performing time-delay sparse filtering on the multi-agent input signal pulse sequence based on the multi-agent signal sparse matrix with time-delay characteristics to obtain a multi-agent time-delay sparse filtered pulse sequence; Step S152: performing rate sparse modulation on the multi-agent time-delay sparse filtered pulse sequence in spatial position distribution based on the multi-agent signal sparse matrix with rate characteristics to generate a multi-agent rate sparse modulated pulse sequence; Step S153: Obtain the pulse energy corresponding to each pulse signal through the multi-agent rate sparse modulation pulse sequence, and perform pulse attenuation evaluation based on the pulse energy corresponding to each pulse signal to obtain the pulse attenuation factor corresponding to the multi-agent pulse signal; Step S154: Based on the pulse attenuation factor corresponding to the multi-agent pulse signal, the spatiotemporal pulse distribution of each corresponding pulse signal in the multi-agent rate sparse modulation pulse sequence is optimized to generate a multi-agent spatiotemporal sparse pulse sequence.

6. The multi-agent privacy protection analysis method for the pulse communication mechanism according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: Constructing a corresponding hierarchical spatiotemporal spike neural network by adopting a structure combining a recursive neural network and a long short-term memory network, specifically including a bottom layer, a middle layer, and a top layer, wherein the bottom layer is used to process the original pulse sequence, the middle layer is used to extract the corresponding pulse timing features, and the top layer is used to decide and output the corresponding communication pulse sequence; Step S22: Input the multi-agent spatiotemporal sparse pulse sequence into the bottom layer of the hierarchical spatiotemporal spiking neural network for pulse preprocessing, and input the preprocessed multi-agent spatiotemporal sparse pulse sequence into the middle layer of the hierarchical spatiotemporal spiking neural network to train the pulse extraction network by introducing a pulse timing attention mechanism. At the same time, a differentially private spike network training method is designed and constructed during the training process to introduce Gaussian noise into the spike neural network and combine the gradient clipping strategy of the training process to prevent gradient reversal attacks, so as to train and generate the pulse timing feature output corresponding to each agent; Step S23: The pulse interaction frequency and pulse sensitivity between each agent are calculated by outputting the pulse timing characteristics corresponding to each agent at the top layer of the hierarchical spatiotemporal pulse neural network, and the corresponding privacy budget is dynamically configured according to the pulse interaction frequency and pulse sensitivity between each agent. parameters, and according to the privacy budget The parameters determine the corresponding communication pulse sequence output in the top layer, generating the multi-agent timing-critical communication pulse sequence.

7. The multi-agent privacy protection analysis method for the pulse communication mechanism according to claim 1 is characterized in that: Step S3 includes the following steps: Step S31: performing pulse timing disturbance analysis on the multi-agent timing-critical communication pulse sequence to obtain the multi-agent communication pulse timing disturbance distribution characteristics; Step S32: Obtain the pulse communication disturbance delay and pulse communication disturbance frequency difference between each agent through the multi-agent communication pulse timing disturbance distribution characteristics, and perform pulse time jitter amplitude statistics on each pulse signal in the multi-agent timing critical communication pulse sequence based on the pulse communication disturbance delay and pulse communication disturbance frequency difference between each agent to obtain the communication pulse signal time jitter amplitude between each agent; Step S33: Generate a corresponding multi-agent pulse time jitter amplitude spectrum based on the time jitter amplitude of the communication pulse signals between each agent, and perform pulse selective deletion optimization on each pulse signal in the multi-agent timing-critical communication pulse sequence based on the multi-agent pulse time jitter amplitude spectrum to filter out redundant pulse signals corresponding to the consistent time jitter amplitude of the communication pulse signals, thereby generating a multi-agent pulse optimized deletion sequence; Step S34: Perform pulse perturbation distribution randomization processing on the timing of the corresponding remaining pulse signals in the multi-agent pulse optimization deletion sequence to generate a multi-agent pulse perturbation randomization sequence.

8. The multi-agent privacy protection analysis method for the pulse communication mechanism according to claim 1 is characterized in that: Step S4 includes the following steps: Step S41: constructing privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, membership inference attack, and attribute inference attack; Step S42: Monte Carlo simulation is used in combination with the corresponding privacy leakage attack scenario model to simulate the privacy leakage attack on the randomized sequence of multi-agent pulse perturbations to generate the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios; Step S43: Based on the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios, the privacy leakage diffusion impact analysis is performed on the multi-agent pulse perturbation randomized sequence to obtain the corresponding diffusion path and impact range of the corresponding pulse privacy leakage event in the multi-agent system; Step S44: Constructing a privacy-efficiency optimization target corresponding to the privacy protection level, communication bandwidth consumption, and computational delay based on the diffusion path and impact range of the corresponding pulse privacy leakage event in the multi-agent system; Step S45: Based on the privacy-efficiency optimization objectives corresponding to the privacy protection level, communication bandwidth consumption and computing delay, a privacy leakage protection analysis is performed on the randomized sequence of multi-agent pulse perturbations, and the Pareto optimal solution set corresponding to the privacy-efficiency optimization objectives is solved by the NSGA-II algorithm. Based on the Pareto optimal solution set, a balance strategy between privacy and efficiency corresponding to the pulse signal communication is obtained to generate a privacy-efficiency protection optimization strategy corresponding to the multi-agent pulse communication.

9. The multi-agent privacy protection analysis method for the pulse communication mechanism according to claim 8, characterized in that: Step S42 includes the following steps: Monte Carlo simulation is used to combine the privacy leakage attack scenario models corresponding to pulse sequence reconstruction attack, membership inference attack and attribute inference attack to simulate and infer the pulse attack of the randomized multi-agent pulse perturbation sequence, so as to generate the corresponding multi-agent pulse perturbation attack generation process under different attack scenarios. Perform multiple rounds of reverse reasoning on the pulse decision-making process of the multi-agent pulse perturbation attack generation process under different attack scenarios to generate the multi-agent pulse perturbation attack decision tree corresponding to different attack scenarios; The corresponding multi-agent pulse perturbation stagnation points in different attack scenarios are obtained through the multi-agent pulse perturbation attack decision trees in different attack scenarios, and the corresponding multi-agent pulse perturbation stagnation points in different attack scenarios are used to predict the leakage probability of the multi-agent pulse perturbation attack decision trees in different attack scenarios to generate the corresponding multi-agent pulse privacy leakage probability in different attack scenarios.

10. A multi-agent privacy protection analysis system based on pulse communication mechanism, characterized in that: A method for performing a multi-agent privacy protection analysis method for an impulse communication mechanism according to claim 1, wherein the multi-agent privacy protection analysis system for the impulse communication mechanism comprises: A spatiotemporal sparse pulse conversion module is used to obtain a continuous input signal sequence corresponding to the multi-agent system and perform spatiotemporal sparse pulse conversion on the continuous input signal sequence corresponding to the multi-agent system based on a sparse matrix generation method of time coding and rate coding to generate a multi-agent spatiotemporal sparse pulse sequence; The pulse communication decision output module is used to construct a hierarchical spatiotemporal pulse neural network and input the multi-agent spatiotemporal sparse pulse sequence into the hierarchical spatiotemporal pulse neural network. Combined with the pulse network training method based on differential privacy, the pulse communication decision output is performed to generate the multi-agent time-critical communication pulse sequence; The pulse perturbation randomization module is used to perform pulse perturbation randomization processing on the multi-agent timing-critical communication pulse sequence based on the combination of pulse time jitter and pulse deletion, thereby generating a multi-agent pulse perturbation randomized sequence; The pulse privacy protection analysis module is used to construct privacy leakage attack scenario models corresponding to pulse sequence reconstruction attacks, member inference attacks, and attribute inference attacks, and uses Monte Carlo simulation combined with the corresponding privacy leakage attack scenario models to simulate privacy leakage attacks on multi-agent pulse perturbation randomized sequences to generate the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios; based on the corresponding multi-agent pulse privacy leakage probabilities under different attack scenarios, the privacy leakage protection analysis of the multi-agent pulse perturbation randomized sequences is performed to generate the privacy-efficiency protection optimization strategy corresponding to multi-agent pulse communication.

Images