A multi-agent privacy protection method based on pulse communication mechanism

Abstract

This invention discloses a multi-agent privacy protection method based on a pulse communication mechanism, relating to the field of artificial intelligence security technology. The method includes: constructing a multi-agent system, and based on the multi-agent system, enabling each agent to synchronously encrypt and share information; wherein the multi-agent system includes multiple agents interconnected via a communication network, and each agent is equipped with: a local policy network, a pulse communication module, and a privacy-utility balance controller; the pulse communication module includes a pulse encoder, a pulse obfuscator, and a pulse decoder. This invention aims to transform the information exchanged between agents from a high-dimensional continuous value vector into a sparse, asynchronous, binary pulse event sequence, reducing the risk of data leakage from an information theory perspective, and naturally compatible with lightweight encryption or obfuscation methods, achieving a good balance between communication efficiency and privacy strength.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence security technology, and in particular to a multi-agent privacy protection method based on pulse communication mechanism. Background Technology

[0002] With the rapid development of the Internet of Things (IoT), edge computing, and distributed artificial intelligence (AI), multi-agent systems have shown great potential in collaborative perception, decision-making, and learning tasks. In such systems, multiple agents (such as mobile devices, sensors, and robots) need to collaborate to achieve a common goal, typically involving joint model training or frequent parameter exchanges. However, traditional collaborative methods based on plaintext gradients, parameters, or state information pose significant privacy risks. Adversaries could intercept or analyze intermediate information exchanged between agents to infer their original local training data, thereby violating user privacy.

[0003] Existing privacy protection techniques, such as differential privacy and secure aggregation in federated learning, offer some protection but still have limitations. Differential privacy can impair model utility by adding noise; methods based on homomorphic encryption or secure multi-party computation incur significant communication and computational overhead, making them unsuitable for resource-constrained agents. Furthermore, traditional continuous-valued communication (such as gradient vectors) has high information density, making it more vulnerable to attacks such as model inversion and member inference.

[0004] Therefore, there is an urgent need for a new mechanism that can enhance privacy in terms of the nature of communication content while achieving efficient collaboration. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-agent privacy protection method based on pulse communication mechanism, which aims to solve or improve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following solution: A multi-agent privacy protection method based on pulse communication mechanism includes: A multi-agent system is constructed, and based on the multi-agent system, each agent synchronously performs information encryption and sharing; wherein, the multi-agent system includes multiple agents interconnected through a communication network, and each agent is also equipped with: a local policy network, a pulse communication module, and a privacy-utility balance controller; the pulse communication module includes a pulse encoder, a pulse obfuscator, and a pulse decoder; The local policy network is used to perform information encryption and sharing computation; the pulse communication module is used to perform encoding, obfuscation and decoding sharing processing on the collaborative information generated in the network; the privacy-utility balance controller is used to dynamically adjust the privacy protection strength of information in the network.

[0007] Optionally, in the multi-agent system, the following processing is performed: First, the cooperative information to be sent is converted into a binary pulse sequence using a pulse encoder. Then, the binary pulse sequence is obfuscated using a pulse obfuscator. The obfuscated pulse sequence is sent to the target receiving agent. The receiving agent decodes the received pulse sequence into a continuous signal using a pulse decoder to recover the information.

[0008] Optionally, the step of converting the coordination information to be transmitted into a binary pulse sequence using a pulse encoder specifically includes: Each agent constructs shared collaborative information based on private data, and then inputs the collaborative information into a pulse encoder to convert the continuous vector signal of the collaborative information into a binary pulse sequence sparsely distributed on the time axis, wherein the pulse firing time and frequency carry the main information.

[0009] Optionally, the step of obfuscating the binary pulse sequence using a pulse obfuscator and sending the obfuscated pulse sequence to the target receiving agent specifically includes: The binary pulse sequence is input into a pulse obfuscator for obfuscation to obtain an obfuscated pulse sequence. The obfuscation process includes: introducing time jitter based on a pseudo-random sequence at the pulse emission time, randomly inserting and discarding pulses according to a set probability, and encrypting the pulse interval. The agent is instructed to send the conflated pulse sequence as an event stream to a designated neighbor agent or central aggregator via communication; the communication is triggered only when a pulse event occurs.

[0010] Optionally, the receiving agent decodes the received pulse sequence into a continuous signal using a pulse decoder to recover the information, specifically including: The receiving agent collects pulse streams and uses a pulse decoder to decode the received pulse sequence into a continuous signal, obtaining either a single original information or aggregated information obtained from multiple pulse streams, thus completing information recovery.

[0011] Optionally, the encoding process of the pulse encoder adopts a bio-inspired encoding strategy, specifically: Each vector dimension of the collaborative information is mapped to an independent leaky integral firing neuron. The membrane potential of each neuron is integrated with the input current of the corresponding dimension over time. When the membrane potential exceeds the threshold, the neuron fires a pulse corresponding to the current, and then the membrane potential is reset and enters the refractory period.

[0012] Optionally, the confusion process of the pulse confused device specifically includes: A time jitter based on a pseudo-random sequence is introduced at the pulse emission time: an offset generated by a private pseudo-random number generator is superimposed on the emission time of each real pulse. The value range of the offset is a privacy hyperparameter, which is eliminated in subsequent data recovery by using a synchronized PRNG and the same seed with the receiver. Random insertion and discarding according to a set probability: Virtual pulses without any information are randomly inserted according to a set insertion probability, or real pulses are randomly discarded according to a set discard probability. In subsequent data recovery, recovery is performed by sharing a random mask derived from a key; the random mask indicates the corresponding virtual pulse or discard position. Encrypted Pulse Interval: Lightweight stream cipher encryption is applied to the pulse interval. In subsequent data recovery, the sender and receiver share the key stream. The sender XORs the original interval with the key stream to obtain the encrypted interval.

[0013] Optionally, the decoding process of the pulse decoder is specifically as follows: For a single information transmission scenario, the pulse stream is decrypted based on the corresponding shared key; For scenarios involving aggregated information transmission, each pulse stream is first deobfuscated according to the corresponding shared key, and then all the deobfuscated pulse streams are superimposed.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses a multi-agent privacy protection method based on a pulse communication mechanism. The method includes constructing a multi-agent system and, based on the multi-agent system, enabling each agent to synchronously encrypt and share information. The multi-agent system includes multiple agents interconnected via a communication network, and each agent is equipped with: a local policy network, a pulse communication module, and a privacy-utility balance controller. The pulse communication module includes a pulse encoder, a pulse obfuscator, and a pulse decoder. The local policy network is used to perform information encryption and sharing computations. The pulse communication module is used to perform encoding, obfuscation, and decoding sharing processes on collaborative information generated in the network. The privacy-utility balance controller is used to dynamically adjust the privacy protection strength of information in the network. This invention aims to transform the information exchanged between agents from a high-dimensional continuous value vector into a sparse, asynchronous, binary pulse event sequence, reducing the risk of data leakage from an information theory perspective, and naturally compatible with lightweight encryption or obfuscation methods, achieving a good balance between communication efficiency and privacy strength. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the structure of the multi-agent system in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The purpose of this invention is to provide a multi-agent privacy protection method based on pulse communication mechanism, which aims to solve or improve at least one of the above-mentioned technical problems.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown, this invention provides a multi-agent privacy protection method based on a pulse communication mechanism, comprising: A multi-agent system is constructed, and based on the multi-agent system, each agent synchronously performs information encryption and sharing; wherein, the multi-agent system includes multiple agents interconnected through a communication network, and each agent is also equipped with: a local policy network, a pulse communication module, and a privacy-utility balance controller; the pulse communication module includes a pulse encoder, a pulse obfuscator, and a pulse decoder.

[0021] The local policy network is used to perform information encryption and sharing computation; the pulse communication module is used to perform encoding, obfuscation and decoding sharing processing on the collaborative information generated in the network; the privacy-utility balance controller is used to dynamically adjust the privacy protection strength of information in the network.

[0022] As a specific implementation method, the following processing is performed in the multi-agent system: First, the cooperative information to be sent is converted into a binary pulse sequence using a pulse encoder. Then, the binary pulse sequence is obfuscated using a pulse obfuscator. The obfuscated pulse sequence is sent to the target receiving agent. The receiving agent decodes the received pulse sequence into a continuous signal using a pulse decoder to recover the information.

[0023] The encoding, obfuscation, communication, and decoding processes described above include: 1. Encoding: Each agent constructs shared collaborative information based on private data, and then inputs the collaborative information into a pulse encoder to convert the continuous vector signal of the collaborative information into a binary pulse sequence sparsely distributed on the time axis, wherein the pulse firing time and frequency carry the main information.

[0024] As a specific implementation method, the encoding process of the pulse encoder adopts a bio-inspired encoding strategy, specifically as follows: Each vector dimension of the collaborative information is mapped to an independent leaky integral firing neuron. The membrane potential of each neuron is integrated with the input current of the corresponding dimension over time. When the membrane potential exceeds the threshold, the neuron fires a pulse corresponding to the current, and then the membrane potential is reset and enters the refractory period.

[0025] 2. Obfuscation: The binary pulse sequence is input into a pulse obfuscator for obfuscation processing to obtain an obfuscated pulse sequence; the obfuscation processing includes: introducing time jitter based on a pseudo-random sequence at the pulse emission time, randomly inserting and discarding according to a set probability, and encrypting the pulse interval.

[0026] As one specific implementation, the confusion process of the pulse confused device specifically includes: A time jitter based on a pseudo-random sequence is introduced at the pulse emission time: an offset generated by a private pseudo-random number generator is superimposed on the emission time of each real pulse. The value range of the offset is a privacy hyperparameter, which is eliminated in subsequent data recovery by using a synchronized PRNG and the same seed with the receiver.

[0027] Random insertion and discarding according to a set probability: Virtual pulses without any information are randomly inserted according to a set insertion probability, or real pulses are randomly discarded according to a set discard probability. In subsequent data recovery, recovery is performed by sharing a random mask derived from a key; the random mask indicates the corresponding virtual pulse or discard position.

[0028] Encrypted Pulse Interval: Lightweight stream cipher encryption is applied to the pulse interval. In subsequent data recovery, the sender and receiver share the key stream. The sender XORs the original interval with the key stream to obtain the encrypted interval.

[0029] 3. Communication: The agent sends the confused pulse sequence as an event stream to a designated neighbor agent or central aggregator via communication; communication is triggered only when a pulse event occurs.

[0030] 4. Decoding: The receiving agent collects pulse streams and uses a pulse decoder to decode the received pulse sequence into a continuous signal, obtaining either a single original information or aggregated information obtained from multiple pulse streams, thus completing information recovery.

[0031] As one specific implementation, the decoding process of the pulse decoder is as follows: For a single information transmission scenario, the pulse stream is decrypted based on the corresponding shared key.

[0032] For scenarios involving aggregated information transmission, each pulse stream is first deobfuscated according to the corresponding shared key, and then all the deobfuscated pulse streams are superimposed.

[0033] Based on the above technical solution, the following embodiments are provided.

[0034] This embodiment uses a multi-agent system consisting of four agents (labeled Agent1, Agent2, Agent3, and Agent4) interconnected via a star-shaped communication network, with Agent1 acting as the central aggregator. Each agent is equipped with a local policy network, a pulse communication module (including a pulse encoder, a pulse obfuscator, and a pulse decoder), and a privacy-utility balance controller. The system aims to achieve secure sharing of collaborative information (such as state observations or gradient updates) among the agents, while balancing privacy and task utility by dynamically adjusting the strength of privacy protection.

[0035] Assume that at the current moment, each agent needs to send its locally computed collaborative information vector (with a dimension of 3) to the central aggregator Agent1. This embodiment uses Agent2 sending information to Agent1 as an example to explain in detail the encoding, obfuscation, communication, and decoding processes of pulse communication.

[0036] Step 1: Encoding.

[0037] Agent2's local policy network generates collaborative information vectors to be shared. X =[ x 1, x 2, x 3], of which x 1 = 0.8 x 2 = -0.3, x3 = 1.2. This vector is input to the pulse encoder, which maps each dimension to an independent leaky integral emission (LIF) neuron, the dynamics of which are described by the following differential equation: in, V i For the first i The membrane potential of a neuron; τ = 10 ms, which is a time constant; input current. I i ( t )= x i (Constant value). When V i Exceed V th At that time, the neuron fires a pulse (output is 1), then... V i Reset to 0 and enter a refractory period of 2ms. The simulation uses a discrete time step Δ. t =1ms, duration T =50ms.

[0038] Encoding process simulation: for x 1 = 0.8, neuron 1 in t =15ms and t A pulse is emitted every 45ms, generating a pulse sequence. S 1 = [0,0,…,1,0,…,1] (only the issuance time is 1).

[0039] for x 2 = -0.3, the input current is negative, the membrane potential remains below the threshold, and no pulse is emitted. S 2 is a sequence of all zeros.

[0040] for x 3 = 1.2, neuron 3 in t =10ms t =32ms and t =48ms pulse output, generating S 3.

[0041] Finally, the pulse encoder outputs a multi-channel binary pulse sequence S=[ S 1, S 2, S 3], each channel corresponds to a dimension, the pulse emission time and frequency (e.g., x The 3 pulses (with higher density) carry information about the original vector.

[0042] Step 2: Obfuscation.

[0043] The pulse obfuscator performs triple obfuscation on S, and the processing parameters are dynamically set by the privacy-utility balance controller according to the current privacy requirements. In this embodiment, the privacy level is set to "medium," and the corresponding parameters are as follows: Time jitter: An offset is added to the output time of each real pulse. The offset is generated by a private pseudo-random number generator (PRNG) with a shared key as the seed. K time The offset range is δ∈[-3ms, 3ms]. For example... S The first pulse of 1 from t =15ms offset to t =17ms.

[0044] Random insertion and discard: Setting the virtual pulse insertion probability p insert =0.2, true pulse drop probability p drop =0.1. Insertion and discard decisions are based on a random mask. M (by key) K mask Derivative) determination, for example in t A dummy pulse is inserted every 5ms, and simultaneously discarded. S 3 in t =32ms real pulse.

[0045] Encryption Pulse Interval: Calculate the interval between adjacent pulses (in milliseconds) using a stream cipher (such as RC4) and a key. K interval Generate a keystream and XOR the intervals. For example, if the original interval value is 8ms and the keystream value is 3, the encrypted interval will be 11ms.

[0046] After obfuscation, the original pulse sequence is converted into an obfuscated sequence, and the pulse firing time, number, and interval are all disturbed, making it difficult to directly infer the original information.

[0047] Step 3: Communication.

[0048] Agent2 treats the obfuscated sequence as an event stream, sending data packets to Agent1 only when a pulse event occurs (i.e., when a "1" appears in the sequence). Each data packet contains the pulse channel number, the obfuscated release time, and the encryption interval. This event-driven communication significantly reduces communication bandwidth; for example, vectors that would otherwise require continuous transmission only need to be sent as small data packets at a few specific moments.

[0049] Step 4: Decode.

[0050] Agent1, acting as the receiver, collects the pulse stream from Agent2 and then uses a pulse decoder to recover the data. Deobfuscation: Agent1 uses the same key synchronized with Agent2. K time , K mask , K interval ) and PRNG, perform the reverse operation: Eliminate time jitter: According to K time Regenerate the offset and reverse the pulse time.

[0051] Restore Insertion and Discard: Based on K mask Identify dummy pulses and their drop locations, remove dummy pulses, and mark the drop locations as missing.

[0052] Decryption interval: Use K interval Generate the same key stream, XOR the encrypted interval to decrypt it, and restore the original interval.

[0053] Information Reconstruction: For a single information transmission scenario (in this example), the deobfuscated pulse sequence is input into the decoder, and a time-based decoding algorithm (such as sliding window averaging) is used to restore the pulse sequence into a continuous value vector. For example, for each channel, the number of pulses is counted within the window [t, t+20ms], and then the cooperative information vector is reconstructed based on the calibration relationship between the pulse frequency and the original value (such as linear mapping).

[0054] Aggregation scenario expansion: If multiple agents send information to Agent1 simultaneously (such as Agent2, Agent3, Agent4), Agent1 can deobfuscate each pulse stream separately, then directly superimpose them in the pulse domain (since pulses are binary events), and then decode to obtain aggregated information, thereby supporting secure distributed aggregation computing.

[0055] Privacy-Utility Balance Control The privacy-utility balance controller dynamically adjusts obfuscation parameters based on task requirements (e.g., high-privacy tasks vs. efficient collaborative tasks). For example, when high privacy protection is required, the time jitter range is adjusted to [-5ms, 5ms], the insertion probability to 0.3, and the discard probability to 0.2; when high communication efficiency is required, the jitter range and insertion / discard probabilities are reduced. The controller can optimize parameters online using reinforcement learning strategies based on historical communication errors and privacy leakage risk assessments.

[0056] Effect verification In this embodiment, the information transmitted by Agent2 to Agent1 is converted into an obfuscated pulse sequence during communication. Even if eavesdropped, it is difficult for an attacker to recover the original data from the pulse event stream, thus protecting the agent's local privacy. Simultaneously, the receiver can accurately recover the information using a shared key, and event-driven communication reduces bandwidth consumption. Experiments show that, under typical parameters, this method can reduce the original information recovery error to less than 5%, while reducing the privacy leakage risk (measured by mutual information) by more than 60%.

[0057] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0058] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multi-agent privacy protection method based on pulse communication mechanism, characterized in that, include: A multi-agent system is constructed, and based on the multi-agent system, each agent synchronously performs information encryption and sharing; wherein, the multi-agent system includes multiple agents interconnected through a communication network, and each agent is also equipped with: a local policy network, a pulse communication module, and a privacy-utility balance controller; the pulse communication module includes a pulse encoder, a pulse obfuscator, and a pulse decoder; The local policy network is used to perform information encryption and sharing computation; the pulse communication module is used to perform encoding, obfuscation and decoding sharing processing on the collaborative information generated in the network; the privacy-utility balance controller is used to dynamically adjust the privacy protection strength of information in the network.

2. The multi-agent privacy protection method based on pulse communication mechanism according to claim 1, characterized in that, In the multi-agent system, the following processing is performed: First, the cooperative information to be sent is converted into a binary pulse sequence using a pulse encoder. Then, the binary pulse sequence is obfuscated using a pulse obfuscator. The obfuscated pulse sequence is sent to the target receiving agent. The receiving agent decodes the received pulse sequence into a continuous signal using a pulse decoder to recover the information.

3. The multi-agent privacy protection method based on pulse communication mechanism according to claim 2, characterized in that, The process of converting the coordination information to be transmitted into a binary pulse sequence using a pulse encoder specifically includes: Each agent constructs shared collaborative information based on private data, and then inputs the collaborative information into a pulse encoder to convert the continuous vector signal of the collaborative information into a binary pulse sequence sparsely distributed on the time axis, wherein the pulse firing time and frequency carry the main information.

4. The multi-agent privacy protection method based on pulse communication mechanism according to claim 2, characterized in that, The step of obfuscating the binary pulse sequence using a pulse obfuscator and sending the obfuscated pulse sequence to the target receiving agent specifically includes: The binary pulse sequence is input into a pulse obfuscator for obfuscation to obtain an obfuscated pulse sequence. The obfuscation process includes: introducing time jitter based on a pseudo-random sequence at the pulse emission time, randomly inserting and discarding pulses according to a set probability, and encrypting the pulse interval. The agent is instructed to send the conflated pulse sequence as an event stream to a designated neighbor agent or central aggregator via communication; the communication is triggered only when a pulse event occurs.

5. The multi-agent privacy protection method based on pulse communication mechanism according to claim 2, characterized in that, The receiving agent decodes the received pulse sequence into a continuous signal using a pulse decoder to recover the information, specifically including: The receiving agent collects pulse streams and uses a pulse decoder to decode the received pulse sequence into a continuous signal, obtaining either a single original information or aggregated information obtained from multiple pulse streams, thus completing information recovery.

6. The multi-agent privacy protection method based on pulse communication mechanism according to claim 1, characterized in that, The encoding process of the pulse encoder adopts a bio-inspired encoding strategy, specifically as follows: Each vector dimension of the collaborative information is mapped to an independent leaky integral firing neuron. The membrane potential of each neuron is integrated with the input current of the corresponding dimension over time. When the membrane potential exceeds the threshold, the neuron fires a pulse corresponding to the current, and then the membrane potential is reset and enters the refractory period.

7. The multi-agent privacy protection method based on pulse communication mechanism according to claim 4, characterized in that, The confusion process of the pulse confusionr specifically includes: A time jitter based on a pseudo-random sequence is introduced at the pulse emission time: an offset generated by a private pseudo-random number generator is superimposed on the emission time of each real pulse. The value range of the offset is a privacy hyperparameter, which is eliminated in subsequent data recovery by using a synchronized PRNG and the same seed with the receiver. Random insertion and discarding according to a set probability: Virtual pulses without any information are randomly inserted according to a set insertion probability, or real pulses are randomly discarded according to a set discard probability. In subsequent data recovery, recovery is performed by sharing a random mask derived from a key; the random mask indicates the corresponding virtual pulse or discard position. Encrypted Pulse Interval: Lightweight stream cipher encryption is applied to the pulse interval. In subsequent data recovery, the sender and receiver share the key stream. The sender XORs the original interval with the key stream to obtain the encrypted interval.

8. The multi-agent privacy protection method based on pulse communication mechanism according to claim 1, characterized in that, The decoding process of the pulse decoder is as follows: For a single information transmission scenario, the pulse stream is decrypted based on the corresponding shared key; For scenarios involving aggregated information transmission, each pulse stream is first deobfuscated according to the corresponding shared key, and then all the deobfuscated pulse streams are superimposed.

Images