Covert communication path construction method and system based on distributed model optimization path

Abstract

The present invention provides a method and system for constructing a covert communication path based on a distributed model optimization path. The method includes: a data sender digitally signs the original data to generate a data signature, constructs embedded information based on the original data and an embedding strategy; encrypts the original data, data signature, and embedded information to generate data ciphertext, and divides it into data blocks according to channel capacity to construct a protocol field; signs the model parameters of the unembedded information to obtain a parameter signature, concatenates the parameter signature, data blocks, and protocol fields to embed the data packet into the model optimization path and sends it; a data receiver extracts the data packet, verifies the parameter signature, and then aggregates it. According to the protocol field, the data blocks are concatenated to obtain data ciphertext, which is then decrypted and verified, ultimately obtaining the original data and the embedding strategy. In this way, communication security is guaranteed while the communication behavior characteristics are concealed, effectively balancing concealment and model performance.

Description

Technical Field

[0001] The present invention relates to the field of data transmission technology, and in particular to a method and system for constructing a covert communication path based on a distributed model optimization path. Background Art

[0002] Amid the rapid development of distributed machine learning, decentralized architectures such as federated learning and cloud-edge collaboration have become key technologies for addressing the "data silo" problem. These technologies are widely used in scenarios such as the Internet of Things (IoT) and smart devices. Each node performs local model training and participates in global model aggregation. For example, in the medical field, hospital nodes serve as both clients and servers, requiring them to protect local data privacy while processing model parameters from other institutions during training.

[0003] However, existing distributed training systems present serious security risks. Attackers can not only conduct model reversal attacks using gradient information, but can also exploit transmitted data such as configuration files and data statistics to infer membership. This is particularly true in scenarios where data is not independent and identically distributed (IID). While traditional encryption methods can ensure data transmission security, they cannot conceal the communication behavior itself, making the transmission process vulnerable to traffic analysis attacks.

[0004] Current communication mechanisms in distributed training face three challenges: first, conventional parameter transmission can easily be identified as a covert communication vector; second, maintaining a stable covert channel is difficult in dynamic network environments; and third, existing methods struggle to balance stealth and model performance. Complex encryption algorithms, especially on resource-constrained edge devices, impose prohibitive computational overhead.

[0005] Existing solutions suffer from the following major drawbacks: First, covert communication based on traffic modulation significantly slows model convergence; second, relying solely on encryption fails to conceal communication characteristics, making them susceptible to detection by deep packet inspection techniques; and finally, static embedding strategies struggle to adapt to dynamically changing network conditions and computational loads. These issues severely limit the effectiveness of covert communication in distributed machine learning.

[0006] In view of the above problems, the existing technology is in urgent need of improvement. Summary of the Invention

[0007] The present invention provides a method and system for constructing a covert communication path based on a distributed model optimization path, which has the advantages of hiding communication behavior characteristics while ensuring communication security and balancing concealment and model performance.

[0008] In one aspect, the present invention provides a method for constructing a covert communication path based on a distributed model optimization path, comprising:

[0009] The data sender digitally signs the original data to generate a data signature, and constructs embedded information based on the original data and the embedding strategy;

[0010] The data sender encrypts the original data, the data signature, and the embedded information to generate data ciphertext, fills the data with a payload according to the channel capacity, divides the data into data blocks, and constructs the protocol fields corresponding to the data blocks;

[0011] The data sender signs the model parameters without embedded information to obtain a parameter signature, and concatenates the parameter signature, the data block, and the protocol field into a data packet, embeds the pre-built model optimization path, and sends the data packet to the data receiver;

[0012] The data receiver extracts the data packet from the model optimization path, verifies the parameter signature to confirm the integrity, and then aggregates the data. After determining that there is a payload in the protocol field, the data block is spliced ​​to obtain the data ciphertext.

[0013] The data receiver decrypts the data ciphertext, decomposes it into the data signature, the original data and the embedded information, verifies the data signature, obtains the original data, and determines the embedding strategy according to the embedded information.

[0014] According to the present invention, a method for constructing a covert communication path based on a distributed model optimization path further includes:

[0015] If the throughput corresponding to the real-time communication demand is greater than or equal to the preset throughput, both the embedding strategy based on the gradient three-dimensional direction and the embedding strategy based on the gradient amplitude are set as the embedding strategy;

[0016] If the throughput corresponding to the real-time communication demand is less than the preset throughput, the embedding strategy based on the gradient three-dimensional direction or the embedding strategy based on the gradient amplitude is set as the embedding strategy.

[0017] According to a method for constructing a covert communication path based on a distributed model optimization path provided by the present invention, the embedding strategy based on the three-dimensional gradient direction is:

[0018] In the first stage, the trained model parameters are recorded, and an optimization trajectory that can map hidden information is constructed in a three-dimensional coordinate system as the model optimization path;

[0019] In the second stage, the direction of the gradient corresponding to the model parameters is controlled so that the model optimization path forms a trajectory with set geometric features in the parameter space, thereby embedding the data packet into the model optimization path.

[0020] According to a method for constructing a covert communication path based on a distributed model optimization path provided by the present invention, the embedding strategy based on gradient amplitude is:

[0021] In the first stage, the first benchmark linear equation is constructed based on the current model parameters and gradient modulus;

[0022] In the second stage, a second benchmark linear equation is constructed based on the next model parameters and the gradient modulus; an initial intersection point of the first benchmark linear equation and the second benchmark linear equation is obtained; if the position of the initial intersection point meets the preset requirement, the initial intersection point is used as the embedding position of the data packet; if the position of the intersection point does not meet the preset requirement, the gradient modulus is dynamically adjusted to correct the equation coefficients; when the position of the intersection point is adjusted to meet the preset requirement, the adjusted intersection point is used as the embedding position of the data packet;

[0023] The preset requirement condition includes that the intersection point is located in the second quadrant or the third quadrant of the rectangular coordinate system.

[0024] According to a method for constructing a covert communication path based on a distributed model optimization path provided by the present invention, before splicing the parameter signature, the data block, and the protocol field into a data packet and embedding it into a pre-constructed model optimization path, the method further includes:

[0025] The data packets are scrambled by using dynamic obfuscation technology to generate data packets in an obfuscated sending queue.

[0026] According to the present invention, a method for constructing a covert communication path based on a distributed model optimization path is provided, which is filled according to the effective load of the channel capacity, comprising:

[0027] The data length of the data ciphertext is extended to an integer multiple of the effective load of the channel capacity.

[0028] According to a method for constructing a covert communication path based on a distributed model optimization path provided by the present invention, the protocol field includes a first field and a second field;

[0029] The first field identifies the load status by the parity of the bit "1", wherein an odd number indicates a valid load and an even number indicates no load;

[0030] The second field is used to record the sequence number of the data block.

[0031] According to a method for constructing a covert communication path based on a distributed model optimization path provided by the present invention, the data signature is generated using the ECDSA algorithm and is obtained by calculating the hash value of the original data using the private key of the data sender;

[0032] The data ciphertext encryption adopts the ECC algorithm and uses the public key of the data receiver to perform the encryption operation.

[0033] According to a method for constructing a covert communication path based on a distributed model optimization path provided by the present invention, the payload of the channel capacity is jointly determined by the model architecture and the parameter quantity; the payload is smaller than the model parameter quantity and includes the total transmission amount of the original data, data signature and protocol information.

[0034] On the other hand, the present invention also provides a covert communication path construction system based on a distributed model optimization path, which includes a data sender and a data receiver;

[0035] The data sender digitally signs the original data to generate a data signature, and constructs embedded information based on the original data and the embedding strategy;

[0036] The data sender encrypts the original data, the data signature, and the embedded information to generate data ciphertext, fills the data with a payload according to the channel capacity, divides the data into data blocks, and constructs the protocol fields corresponding to the data blocks;

[0037] The data sender signs the model parameters without embedded information to obtain a parameter signature, and concatenates the parameter signature, the data block, and the protocol field into a data packet, embeds the pre-built model optimization path, and sends the data packet to the data receiver;

[0038] The data receiver extracts the data packet from the model optimization path, verifies the parameter signature to confirm the integrity, and then aggregates the data. After determining that there is a payload in the protocol field, the data block is spliced ​​to obtain the data ciphertext.

[0039] The data receiver decrypts the data ciphertext, decomposes it into the data signature, the original data and the embedded information, verifies the data signature, obtains the original data, and determines the embedding strategy according to the embedded information.

[0040] The present invention provides a method and system for constructing a covert communication path based on a distributed model optimization path. It integrates technical means such as public-private key pair generation, digital signatures, data encryption, and protocol field parsing, aiming to address data privacy protection in distributed training scenarios and ensure the security, confidentiality, and integrity of data transmission. Encryption and signature technologies ensure the confidentiality and authenticity of data; dynamic obfuscation and model optimization path embedding effectively enhance the concealment of communication; data segmentation and splicing strategies ensure transmission efficiency; and flexible caching strategies and data integrity verification enhance the adaptability and traceability of the system. This solution achieves privacy protection while ensuring computational efficiency, providing an efficient, secure, and reliable covert communication solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 This is a flow chart of a method for constructing a covert communication path based on a distributed model optimization path provided by an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of embedding data based on the embedding strategy of the gradient three-dimensional direction;

[0044] Figure 3 This is another flowchart of a method for constructing a covert communication path based on a distributed model optimization path provided by an embodiment of the present invention;

[0045] Figure 4 It is a structural diagram of a covert communication path construction system based on distributed model optimization path provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0047] Among related technologies, distributed machine learning is widely used in privacy-sensitive fields such as healthcare and finance. However, data such as gradient information and configuration files transmitted during training poses a risk of leakage. While traditional encryption methods can ensure data security, they significantly increase communication overhead and expose transmission intentions. Existing covert communication schemes often rely on fixed embedding strategies, making it difficult to adapt to dynamically changing network environments and differentiated transmission requirements. For example, in the case of hospital federated learning, various institutions need to synchronously transmit patient vital sign data when exchanging model parameters. Traditional methods cannot achieve covert and efficient transmission while ensuring privacy.

[0048] To address these issues, research has found that covert channels must meet three requirements: first, the embedding process must not affect model convergence performance; second, the transmission mechanism must be dynamically adaptable; and third, data integrity must be verifiable. By analyzing the model parameter update trajectory, it was discovered that changes in gradient direction and amplitude can construct an information mapping space. Based on this, a method was proposed to encode sensitive data as geometric features of the optimization path, using digital signatures to ensure end-to-end security and combining dynamic obfuscation techniques to enhance anti-analysis capabilities.

[0049] Therefore, the present invention proposes a covert communication path construction method based on a distributed model optimization path, in which the data sender digitally signs the original data to generate a data signature, and constructs embedded information according to the original data and the embedding strategy; encrypts the original data, data signature and embedded information to generate data ciphertext and divides it into data blocks according to the channel capacity, and constructs a protocol field; signs the model parameters of the unembedded information to obtain a parameter signature, splices the parameter signature, data block and protocol field to embed the data packet into the model optimization path and sends it; the data receiver extracts the data packet to verify the parameter signature and then aggregates it, splices the data blocks according to the protocol field to obtain the data ciphertext and decrypts and verifies it, and finally obtains the original data and the embedding strategy.

[0050] Specifically, Figure 1 The present invention provides a flow chart of a method for constructing a covert communication path based on a distributed model optimization path.

[0051] like Figure 1 As shown, the execution subject of the covert communication path construction method based on distributed model optimization path provided by the embodiment of the present invention can be an electronic device, and the method mainly includes the following steps:

[0052] 101. The data sender digitally signs the original data to generate a data signature, and constructs embedded information based on the original data and the embedding strategy;

[0053] 102. The data sender encrypts the original data, the data signature, and the embedded information to generate data ciphertext, fills the data with a payload according to the channel capacity, divides the data into data blocks, and constructs the protocol fields corresponding to the data blocks;

[0054] 103. The data sender signs the model parameters without embedded information to obtain a parameter signature, and concatenates the parameter signature, the data block, and the protocol field into a data packet, embeds the pre-built model optimization path, and sends the data packet to the data receiver;

[0055] 104. The data receiver extracts the data packet from the model optimization path, verifies the parameter signature to confirm integrity, and then aggregates the data. After determining that there is a payload in the protocol field, the data block is spliced ​​to obtain the data ciphertext.

[0056] 105. The data receiver decrypts the data ciphertext, decomposes it into the data signature, the original data, and the embedded information, verifies the data signature, obtains the original data, and determines the embedding strategy according to the embedded information.

[0057] In a specific implementation, digital signature generation can utilize the Elliptic Curve Digital Signature Algorithm (ECDSA), which uses the sender's private key to encrypt the original data hash value to generate an unforgeable signature, ensuring the authenticity of the data source. Encryption operations utilize the Elliptic Curve Cryptography (ECC) algorithm, asymmetrically encrypting the data using the receiver's public key to achieve information hiding during transmission. The protocol field can include a payload status identifier and a data block sequence number, with the parity bit indicating the presence of the payload. The sequence number field supports reassembly of out-of-order data. Model optimization path construction can be based on parameter update trajectories, forming a resolvable geometric feature path by manipulating the gradient direction or amplitude.

[0058] Specifically, after the data sender generates a digital signature locally, it encapsulates the original data and the embedding strategy into embedded information, where the embedded information is used to set the subsequent caching strategy. After filling the payload according to the channel capacity, the ciphertext is divided into data blocks marked with serial numbers, and the sending order is disrupted through a dynamic obfuscation mechanism. After the receiver extracts the data packet, it first verifies the parameter signature to ensure that the model has not been tampered with, and then parses the protocol field to identify the payload status. When valid data is detected, the data block is reassembled according to the serial number and decrypted to restore the original information. The entire transmission process is deeply coupled with the model training process, and the gradient update path carries both parameter optimization information and maps hidden data. Among them, the channel capacity configuration can be determined by the model architecture and the number of parameters, maximizing the efficiency of information transmission while ensuring stable model training.

[0059] Compared to traditional schemes, which use fixed encryption modes and result in noticeable communication signatures, this embodiment dynamically identifies payload status through protocol fields. During idle periods, only signature parameters are transmitted, enabling model optimization and effectively reducing detectability. When payload is present, data extraction can be performed while model optimization is ongoing. Compared to covert transmission methods that rely solely on gradient modification, this embodiment introduces a dual signature mechanism, implementing integrity verification at both the data and parameter levels to establish a multi-layered security protection system.

[0060] Through the above technical solution, the present invention enables the covert transmission of sensitive data during model parameter updates, resolving the difficult balance between privacy leakage and communication efficiency in traditional federated learning scenarios. The sender dynamically adjusts the embedding strategy to adapt to network conditions, and the receiver quickly identifies the payload through protocol fields, achieving secure data exchange while ensuring rapid model convergence. A dual-signature mechanism effectively protects against man-in-the-middle attacks and data tampering, while dynamic obfuscation technology significantly enhances the untraceability of the transmission path.

[0061] In a specific implementation, the present invention further proposes a method for dynamically configuring an embedding strategy based on a comparison of the throughput corresponding to the real-time communication demand with a preset throughput. When the throughput corresponding to the real-time communication demand is greater than or equal to the preset throughput, both the embedding strategy based on the three-dimensional gradient direction and the embedding strategy based on the gradient magnitude are set as the embedding strategy. When the throughput corresponding to the real-time communication demand is less than the preset throughput, either the embedding strategy based on the three-dimensional gradient direction or the embedding strategy based on the gradient magnitude is selected as the embedding strategy. The specific choice can be set based on actual needs.

[0062] The embedding strategy based on the three-dimensional gradient direction encodes a specific geometric trajectory in the three-dimensional parameter space by adjusting the gradient direction. This is achieved by recording the model parameter update trajectory and constructing an optimized path in the parameter space that maps the hidden information. This strategy embeds information through a two-stage iterative process: the first stage records the original gradient information, and the second stage adjusts the gradient direction to form a preset trajectory. This achieves data embedding without changing the model's convergence direction.

[0063] Gradient amplitude-based embedding refers to an encoding method that modulates the intersection coordinates of linear equations by adjusting the gradient modulus. This is achieved by constructing a baseline linear equation and a comparison linear equation, then adjusting the gradient modulus coefficients to adjust the equation coefficients so that the intersection coordinates meet the preset quadrant conditions. This strategy uses a two-stage adjustment process: the first stage establishes the baseline equation, and the second stage dynamically adjusts the gradient modulus to change the intersection position, mapping binary information to the coordinate quadrant difference.

[0064] Specifically, in high-throughput scenarios, the two embedding strategies can be executed in parallel. For example, when the channel bandwidth permits, the gradient direction adjustment module and the amplitude control module can work simultaneously, respectively processing parameter updates in different dimensions. The direction strategy constructs a spiral trajectory in three-dimensional space, and the amplitude strategy generates a broken line trajectory in a two-dimensional plane. The two trajectories are superimposed to form a composite coding mode, which allows double the amount of data to be transmitted per unit time. In low-throughput scenarios, the system automatically selects a single strategy based on the current resource status. For example, when computing resources are limited, the amplitude strategy is preferred because it only needs to adjust the gradient modulus and has a lower computational complexity; when the communication quality is unstable, the direction strategy is selected because it is more robust to parameter changes.

[0065] Compared to traditional covert communication methods, which typically use fixed coding strategies and are unable to adapt to dynamically changing network environments, traditional solutions can waste computing resources when bandwidth is sufficient, but struggle to guarantee transmission efficiency when resources are limited. This embodiment achieves an intelligent balance between transmission capacity and computational load through a dynamic switching mechanism between two strategies. This results in higher transmission efficiency in high-load scenarios and lower computational overhead in low-load scenarios, while maintaining a significant impact on model convergence speed.

[0066] Through the above-mentioned technical solution, the present invention effectively resolves the contradiction between transmission efficiency and resource consumption in covert communication scenarios. In scenarios such as joint medical imaging training, which require the transmission of large amounts of diagnostic data, the system automatically activates dual-strategy mode, utilizing a dual encoding mechanism of three-dimensional trajectories and two-dimensional coordinates to achieve high-speed transmission. In resource-constrained scenarios such as collaborative learning of smart home devices, it switches to single-strategy mode, ensuring real-time performance by streamlining the computational process. This dynamic adaptation mechanism enables the covert communication system to maintain optimal operating conditions in different application scenarios, balancing transmission security and operational efficiency.

[0067] In a specific implementation process, the present invention further proposes an embedding strategy based on the three-dimensional direction of the gradient. In the first stage, the trained model parameters are recorded and an optimization trajectory that can map hidden information is constructed in the three-dimensional coordinate system as the model optimization path; in the second stage, the direction of the gradient corresponding to the model parameters is controlled so that the model optimization path forms a trajectory with set geometric features in the parameter space, thereby realizing the embedding of the data packet into the model optimization path.

[0068] Among them, the optimization trajectory in the three-dimensional coordinate system refers to the continuous path formed in the three-dimensional parameter space by recording the update sequence of the weight parameters during the model training process. Specifically, it can be visualized by using the projection coordinates of the parameter vector in three orthogonal dimensions. The trajectory can reflect the dynamic evolution of the model parameters. Gradient direction control refers to adjusting the spatial orientation of the gradient vector during the parameter update process. Specifically, it can be achieved by constraining the projection angle of the gradient vector in the selected three-dimensional subspace, so that the optimization path forms a preset angle or intersection coordinates and other geometric features in the parameter space. The set geometric feature refers to a pre-defined trajectory morphology identification mark, which can be specifically a parseable geometric figure such as a triangle, a broken line or a spiral line, which is used to encode binary information.

[0069] Specifically, during the model training process, the gradient information is first fully recorded in the regular parameter update phase to form an initial optimization trajectory. Subsequently, in the second phase, the gradient direction is fine-tuned, and by constraining the projection angle of the gradient vector in three-dimensional space, the trajectory segments formed by two consecutive parameter updates form a specific angle. For example, when a bit "1" needs to be embedded, the angle between adjacent trajectory segments is adjusted to an acute angle; when a bit "0" needs to be embedded, it is adjusted to an obtuse angle. The receiver can reversely parse the embedded hidden information by analyzing the differences in the geometric features of the trajectory morphology. This process maintains the effectiveness of the model update and achieves the low-disturbance characteristics of information embedding by directionally controlling the gradient direction rather than directly modifying the parameter value.

[0070] Figure 2 This is a schematic diagram of embedding data based on the embedding strategy of the gradient three-dimensional direction, such as Figure 2 As shown in the figure, the model structure consists of linear and convolutional layers. Three model parameters are selected to construct a parameter space, within which the gradient update trajectory of each iteration is recorded. The figure clearly illustrates the training status of the model at time points t-1, t, and t+1. In the first iteration, from t-1 to t, nodes generate gradients and update weights, forming the first update path. Then, in the second iteration, from t to t+1, nodes generate new gradients and fine-tune weights, constructing the second update path to embed information. In the second iteration, the upper path corresponds to the stegopath generated when embedding a message bit "1," while the lower path represents the stegopath constructed when embedding a message bit "0." This mechanism provides the technical foundation for covert communication. Using the gradient information of the model parameter updates in each iteration, covert data is embedded in the direction or magnitude of the gradient. Its core design ensures that the main direction of the gradient remains unchanged, ensuring that model updates always follow the optimal convergence path. In this way, each node can still effectively participate in the aggregation and update of the global model after embedding hidden information, achieving information transmission while maintaining the integrity of the learning process.

[0071] Compared with traditional covert communication methods, which typically embed information by directly modifying the numerical values ​​of model parameters, this method can easily lead to abnormal model convergence or performance degradation. This embodiment constructs a geometric optimization path, encoding information into differentiated features of trajectory morphology. While maintaining the overall magnitude of the gradient vector, only its directional component is adjusted. This ensures that the model update process continues along the optimal convergence direction, effectively avoiding model distortion caused by parameter tampering.

[0072] Through the above technical solution, the present invention enables reliable transmission of covert information during distributed model training, carrying communication data by optimizing the geometric changes of the path without interfering with normal model convergence. This method leverages the inherent characteristics of model parameter updates to achieve information embedding, deeply coupling the covert communication process with the training process. This ensures the confidentiality and security of data transmission while maintaining the stability and efficiency of model training.

[0073] In a specific implementation process, the present invention further proposes an embedding strategy based on gradient amplitude: in the first stage, a first benchmark linear equation is constructed based on the current model parameters and gradient modulus; in the second stage, a second benchmark linear equation is constructed based on the next model parameters and gradient modulus; the initial intersection of the first benchmark linear equation and the second benchmark linear equation is obtained, and if the position of the initial intersection meets the preset requirement conditions, the initial intersection is used as the embedding position of the data packet; if the position of the intersection does not meet the preset requirement conditions, the gradient modulus is dynamically adjusted to correct the equation coefficients, and when the position of the intersection is adjusted to meet the preset requirement conditions, the adjusted intersection is used as the embedding position of the data packet; wherein, the preset requirement conditions include the intersection being located in the second quadrant or the third quadrant of the rectangular coordinate system.

[0074] Among them, the first benchmark linear equation refers to a linear equation constructed based on the current model parameters and the gradient modulus. Specifically, the equation can be constructed using the product relationship between the parameter vector and the gradient modulus. The first benchmark linear equation can be a1x+b1y-c1=0. The second benchmark linear equation refers to a linear equation constructed based on the updated model parameters and the adjusted gradient modulus. Specifically, the equation can be generated using a linear combination of the vector modulus after the parameter update and the gradient adjustment amount to form a contrasting geometric structure. Specifically, the equation can be a2x+b2y-c2=0. The initial intersection refers to the coordinates of the intersection of the two benchmark linear equations when the gradient modulus is not adjusted. Specifically, the intersection position can be solved by a simultaneous equation to determine whether the preset quadrant condition is met. Dynamically adjusting the gradient modulus refers to adjusting the amplitude of the gradient vector in real time based on the deviation between the intersection position and the preset condition. Specifically, the modulus can be corrected using a gradient scaling factor. By changing the equation coefficient, the intersection is shifted to the target quadrant to ensure the accuracy of information embedding.

[0075] Specifically, during model training, the data sender first constructs a first baseline linear equation based on the current model parameter vector and the gradient norm. This equation reflects the geometric relationship in the current parameter space. During the next parameter update, a second baseline linear equation is constructed based on the updated parameter vector and the original gradient norm. The intersection of these two equations may be in a non-target quadrant. The system calculates the coordinates of the initial intersection point using simultaneous equations. If it is in the second or third quadrant, this point is directly used as the data embedding location. When yi ≥ 0 (i.e., the i-th intersection point is in the second quadrant), it indicates "1," and when yi < 0 (i.e., the i-th intersection point is in the third quadrant), it indicates "0." If it is in other quadrants, the coefficients of the second equation are modified by dynamically adjusting the gradient norm. For example, a scaling factor is applied to the gradient vector, so that the intersection of the modified second baseline linear equation and the first equation is shifted to the target quadrant. This process achieves precise control of the geometric relationship by fine-tuning the gradient amplitude, ensuring the proper update of model parameters while deeply coupling the data embedding process with the gradient adjustment process, effectively concealing communication traces.

[0076] Compared with traditional covert communication methods, which typically embed information solely by adjusting the gradient direction, this method can easily lead to reduced model convergence performance due to directional deviations. This embodiment combines a dual control mechanism of gradient direction and modulus length. While maintaining the stability of the main gradient direction, it uses modulus length fine-tuning to achieve precise geometric structure construction. This reduces interference with the model training process and improves the stealth of information embedding. Furthermore, by presetting the second and third quadrants as the embedding regions, coordinate symbol conflicts are avoided, enhancing the robustness of the receiver's recognition of the embedded location.

[0077] Through the above-mentioned technical solution, the present invention solves the existing problem of model performance degradation caused by excessive gradient adjustment, achieving covert data embedding without affecting the model's convergence speed. A dynamic gradient modulus length adjustment mechanism ensures that data packets are always embedded in the preset quadrant, significantly improving the accuracy of information extraction and anti-interference capabilities. Furthermore, by combining the data embedding process with the geometric structure of the parameter space, the risk of covert communication behavior being detected as an anomaly is effectively reduced.

[0078] In a specific implementation process, the present invention further proposes to scramble the data packets through dynamic obfuscation technology to generate data packets for the obfuscated sending queue before splicing the parameter signatures, data blocks and protocol fields into data packets and embedding them into a pre-built model optimization path.

[0079] Dynamic obfuscation technology involves changing the transmission order of data packets through a randomization algorithm. This can be achieved using a pseudo-random permutation algorithm or a hash chain-based sequence generation method. This disrupts the original order of data packets, preventing external attackers from inferring the true transmission logic of the packets through traffic analysis. The obfuscated send queue refers to the collection of data packets to be sent after out-of-order processing. This can be achieved by combining timestamps with random numbers to generate a dynamic send sequence. This hides the correlation between data blocks, making it difficult for attackers to reconstruct the complete information flow by intercepting partial packets.

[0080] Specifically, before sending a data packet, the sender inputs the encapsulated data packet sequence into the dynamic obfuscation module, which rearranges the packets according to a preset randomization strategy. For example, a hash chain-based obfuscation algorithm is used to iteratively calculate the mixed result of the current timestamp and the packet hash value to generate a dynamic sorting index, converting the original ordered queue into an unordered transmission sequence. After extracting the data packet, the receiver parses the sequence number information in the protocol field and reassembles the data blocks in the original order to ensure accurate data restoration. This entire process does not change the content structure of the data packet, only adjusting its transmission order. This effectively improves the anti-analysis capabilities of covert communications without adding additional communication overhead.

[0081] Compared to traditional covert communication methods, which typically transmit packets in a fixed order, attackers can infer communication behavior by analyzing packet arrival times or traffic patterns. Dynamic obfuscation, on the other hand, introduces a randomization mechanism that makes the order in which each packet is sent independent of its content logic. Even if an attacker intercepts all packets, it is difficult to restore the original data stream through timing analysis or pattern matching, significantly enhancing the communication system's anti-tracing and anti-interference capabilities.

[0082] Through the above technical solution, the present invention effectively solves the problem that the data packet transmission mode is easily identifiable in traditional covert communications. The dynamic obfuscation mechanism disrupts the sending order of data packets, making it impossible for external observers to infer the actual data transmission logic through traffic analysis, thereby maintaining communication efficiency while improving the security and reliability of covert information transmission.

[0083] It should be noted that the above-mentioned disordering of the data packets may specifically be disordering of the data blocks forming the data packets.

[0084] Specifically, the data sender fills the data ciphertext to an integer multiple of the channel capacity's payload, then divides it into several labeled data blocks, and finally implements dynamic obfuscation sending through dynamic obfuscation technology, thereby significantly improving the stealth and anti-tracking capabilities of the communication.

[0085] Ciphertext data padding refers to extending the encrypted data length to an integer multiple of the channel capacity payload. Specifically, this can be achieved using PKCS#7 padding to ensure that the size of the segmented data blocks matches the channel capacity, avoiding transmission anomalies caused by insufficient data length. Data block labeling refers to assigning a unique sequence identifier to each segmented data block. This can be achieved using an incremental number or hash value generation algorithm, allowing the recipient to reassemble the data based on the sequence number index in the protocol field.

[0086] Specifically, after encrypting and generating the data ciphertext, the data sender first determines the integer multiple length of the payload based on the channel capacity and performs a tail padding operation on the data ciphertext. After the padding is completed, the data ciphertext is divided into multiple labeled data blocks of fixed length, and each data block carries a unique serial number identifier. Subsequently, dynamic obfuscation technology is used to shuffle the data blocks to generate a randomly arranged sending queue. During the transmission process, the data blocks are embedded in the model optimization path in a non-continuous and nonlinear order. The receiver needs to reversely reassemble according to the serial number information in the protocol field to restore the original data ciphertext. For example, data blocks may be sent in the order of 3-1-4-2, rather than the original order of 1-2-3-4. This randomized transmission method makes it difficult for external observers to infer the correlation between data blocks through traffic analysis.

[0087] Compared to traditional covert communication methods, which typically transmit data blocks in a fixed order, attackers can recover the original information by analyzing the temporal correlation or positional continuity of packets. This embodiment, however, introduces a dynamic obfuscation mechanism to randomize the transmission order at the data block level. This eliminates predictable correlations between adjacent packets, effectively blocking attacks based on traffic pattern analysis.

[0088] Through the above technical solutions, the present invention significantly enhances the anti-tracking capabilities and security of covert communications. Dynamic obfuscation technology increases the randomness of data block distribution, making it impossible for attackers to infer the entire transmission content by intercepting partial data blocks. At the same time, the coordinated design of labeled data blocks and protocol fields ensures that the receiver can accurately restore the original data. This ensures high confidentiality and anti-interference capabilities of the communication process while ensuring transmission reliability.

[0089] In a specific implementation process, the present invention further proposes that the protocol field includes a first field and a second field, the first field identifies the load status through the parity of the bit "1", where odd numbers represent valid loads and even numbers represent no loads, and the second field is used to record the serial number of the data block.

[0090] Among them, the first field refers to an identification field that characterizes whether the data packet carries valid data by counting the parity of the number of "1"s in the bits. Specifically, it can be implemented using a 4-bit binary number. The data extraction mechanism of the receiver is triggered by judging whether the total number of "1"s in the first 4 bits is odd or even. This design can quickly identify the payload status with extremely low computing overhead and avoid complex decoding operations.

[0091] Among them, the second field refers to the index field used to mark the position order of the data block in the original data ciphertext. It can be implemented using an 8-bit binary number. By recording the serial number in the range of 0-255, the receiver can accurately reassemble the data block in the disordered transmission scenario. This mechanism effectively solves the problem of data packet disorder caused by dynamic obfuscation transmission.

[0092] Specifically, when constructing the protocol field, the data sender first generates a sequence number based on the number of data blocks after the data ciphertext is split, and encodes it into the 8-bit space of the second field. The first 4 bits of the first field are initialized to a specific bit combination, and the parity is adjusted to match the payload status of the current data packet by adjusting the number of "1"s. For example, when the data packet carries a valid data block, the first 4 bits are set to a bit combination containing an odd number of "1"s; if it is in an empty state, it is set to an even number of "1"s. After extracting the protocol field, the data receiver first analyzes the parity of the first field to determine whether to start the data reassembly process. If a valid payload is detected, the data blocks that arrive out of order are sorted according to the sequence number recorded in the second field, and finally the complete ciphertext data is restored. This process realizes the coordinated operation of transmission status identification and data block sequence management through the dual control function of the protocol field.

[0093] Compared to traditional covert communication methods, which typically use a fixed-length header to identify the payload status, this approach lacks a dynamic sequence management mechanism, making it difficult for the receiver to accurately reconstruct data in out-of-order transmission scenarios. In existing solutions, data block ordering relies on transmission timing or the addition of additional timestamps, increasing protocol overhead and computational complexity. This embodiment integrates the payload status identifier and sequence index into a fixed-length protocol field. This allows for accurate reconstructing of out-of-order data blocks while maintaining low communication overhead, resolving the data misordering issue caused by dynamically obfuscated transmissions.

[0094] Through the above-mentioned technical solution, the present invention significantly improves the efficiency of data block reassembly during covert communication, reducing the risk of packet loss or mis-splicing due to out-of-order transmission. The dual-field structure of the protocol field allows the receiver to parse only 12 bits of information to determine the payload status and extract the sequence number. Compared to traditional multi-field separation designs, this reduces data parsing time and computing resource consumption. This mechanism effectively improves the reliability and integrity of data transmission while ensuring confidentiality.

[0095] In a specific implementation process, the present invention further proposes that the data signature is generated using the ECDSA algorithm, which is obtained by calculating the hash value of the original data using the private key of the data sender; the data ciphertext is encrypted using the ECC algorithm, and the encryption operation is performed using the public key of the data receiver.

[0096] Among them, the ECDSA algorithm refers to a digital signature algorithm based on elliptic curve cryptography, which can be implemented using the secp256k1 elliptic curve parameters. The hash value of the original data is encrypted with the private key to generate a digital signature, which is used to verify the integrity of the data and the authenticity of the sender's identity.

[0097] Among them, the ECC algorithm refers to the elliptic curve encryption algorithm, which can be implemented using an elliptic curve with a 256-bit key length. The original data, data signature and embedded information are encrypted using the public key of the data recipient to generate irreversible ciphertext, which is used to ensure the confidentiality and anti-cracking capabilities of the transmitted data.

[0098] Specifically, during the preprocessing phase, the data sender performs a SHA-256 hash operation on the original data to generate a digest. The sender then uses its private key to encrypt the digest using the ECDSA algorithm to generate a digital signature. This signature, along with the original data and embedded policy parameters, forms the encrypted data unit, which is then encrypted using the receiver's public key using ECC to generate structured ciphertext. During the decryption phase, the receiver uses the corresponding private key to restore the ciphertext, extracts the digital signature, and verifies it using the sender's public key. By comparing the hash values, the receiver confirms that the data has not been tampered with during transmission.

[0099] Compared with traditional schemes, which often use the RSA algorithm for signing and encryption, the key length must reach 3072 bits to achieve the same security strength as 256-bit ECC, resulting in increased computational load and reduced communication efficiency. This embodiment significantly reduces the key length by using elliptic curve cryptography while maintaining the same security level, speeding up signature generation by approximately 40% and reducing the encrypted data size by 60%, making it more suitable for resource-constrained distributed covert communication scenarios.

[0100] Through the above technical solutions, the present invention achieves end-to-end data integrity protection and resistance to man-in-the-middle attacks, ensuring the security of covert communications while optimizing computing resource utilization. The ECDSA signature mechanism effectively prevents data forgery and tampering, while the ECC encryption algorithm maintains efficient transmission even under low-bandwidth conditions. This dual technology combination solves the problem of excessive computational overhead in traditional encryption methods in distributed environments.

[0101] In a specific implementation process, the present invention further splices parameter signatures, data blocks and protocol fields into data packets and embeds them into a pre-built model optimization path. The iterative update of model parameters constructs a dynamic optimization path, which visualizes the evolution trajectory of weight parameters during training and realizes directional mapping of data information representation space through parameter tuning.

[0102] The dynamic optimization path refers to the update trajectory of the weight parameters formed over multiple iterations during model training. This can be achieved by using a gradient descent algorithm to generate a continuous moving trajectory in parameter space. This trajectory is visualized by recording the coordinate changes of each parameter update. Parameter tuning refers to fine-tuning the model weights to change the geometry of the optimization path. This can be achieved by adjusting the gradient direction or amplitude to modify the parameter update direction. This adjustment causes the optimization path to form a specific geometric structure in parameter space, thereby encoding the data packet as a changing pattern of the path shape.

[0103] Specifically, during the model training process, the data sender continuously records the updated coordinates of the weight parameters and constructs an optimization trajectory map in three-dimensional space. When it is necessary to embed a data packet, the sender selectively adjusts the gradient direction or amplitude based on the binary sequence of the information to be transmitted, so that the trajectory formed by the subsequent parameter update deviates from the original path. For example, when it is necessary to embed a bit "1", the sender increases the parameter update amplitude by increasing the gradient modulus, causing the optimization path to form a line segment with a steep slope in three-dimensional space; when it is necessary to embed a bit "0", the gradient modulus is reduced to generate a gentle path change. The receiver continuously monitors the differences in the geometric features of the parameter update trajectory and reversely parses the embedded binary data sequence. During this process, the parameter tuning operation is limited to the preset perturbation threshold to ensure that the model convergence performance is not significantly affected.

[0104] In some embodiments, the dynamic optimization path can be constructed by selecting three key weight parameters as coordinate axes, and recording their numerical changes after each iteration to form three-dimensional trajectory points. For example, in a fully connected layer of a neural network, the weight values ​​of three adjacent neurons are selected as coordinate references. By adjusting the gradient update amplitude of these weights, the optimization path forms a spiral or broken line trajectory in three-dimensional space. Different trajectory shapes correspond to different data encoding rules.

[0105] Compared to traditional covert communication methods, which typically rely on fixed transmission paths or predefined encoding rules, this approach is easily identified by traffic analysis tools. This embodiment, however, leverages the parameter update trajectories naturally generated during model training as dynamic carriers, fine-tuning gradients to achieve directional changes in path morphology. This dynamic path construction method fully integrates the data embedding process with model training, making it difficult for external observers to distinguish between normal parameter updates and covert information transmission, significantly improving the stealth of communication.

[0106] Through the above technical solution, the present invention solves the problems of existing covert communication methods, such as fixed paths that are easily detected and static encoding rules that are easily cracked. Dynamically optimizing the path adaptively changes as the model training progresses, allowing for continuous updates of data embedding locations and encoding rules, effectively defending against detection attacks based on traffic pattern analysis. Furthermore, parameter tuning operations are deeply coupled with the gradient update process, enabling covert data transmission while maintaining the continuity and convergence efficiency of model training.

[0107] In a specific implementation process, the overall implementation process of the covert communication path construction method based on distributed model optimization path of the present invention can be found in Figure 3 . Figure 3 FIG. 1 is another flow chart of a method for constructing a covert communication path based on a distributed model optimization path provided by an embodiment of the present invention. Figure 3 As shown, the data sender and the data receiver will first generate a public-private key pair locally and make the public key public, deploy the model at the same time, and transmit the data request, that is, step 501; the data sender first digitally signs the original data to generate a data signature, that is, step 301; the data sender constructs the embedding information according to the length of the original data and the subsequent embedding strategy, that is, step 302; the data sender uses the public key of the data receiver to encrypt the original data, the data signature generated in step 301, and the embedding information constructed in step 302 to generate a data ciphertext, that is, step 303; the data sender fills the data ciphertext generated in step 303 to an integer multiple of the channel capacity payload, and then divides it into several labeled data blocks, that is, step 304; the data sender uses its own private key to encrypt the unadjusted embedded The model parameters of the message are digitally signed to generate a parameter signature, i.e., step 305; the data sender constructs a 12-bit protocol field, the first 4 bits use the parity of "1" to identify the load status (odd valid / even no-load), and the last 8 bits record the data block sequence number (0-255), realizing the state recognition and disorderly reorganization function, i.e., step 306; the data sender encapsulates the data block of step 304, the digital signature generated in step 305, and the protocol field constructed in step 306 into a data packet, i.e., step 307; the initial state of the data packet sending queue is ordered, and a confused sending queue is generated by dynamic obfuscation technology, i.e., step 308; the data sender converts the data packet in the confused sending queue generated in step 308 into binary, i.e., step 309; the data sender uses Figure 2 As shown, the binary data package is embedded into the model optimization path, i.e., step 310.

[0108] The data sender and the data receiver respectively send and receive the model parameters through 502. Figure 2As shown, a data packet is extracted from the optimized path, i.e., step 401; the data receiver analyzes the protocol field, i.e., step 402; the data receiver determines whether the payload state carries data based on the protocol field in step 402, i.e., step 403; the data receiver uses the public key of the data sender to verify the parameter signature of the model parameter, i.e., step 404; if 404 verification passes and 403 determines that data is carried, the data block is extracted from the data packet, i.e., step 405; the data receiver uses the serial number of the protocol field to splice the data blocks to form a data ciphertext, i.e., step 406; the data receiver decrypts the data ciphertext reconstructed in step 406 using its own private key and decomposes it into embedded information, original data and data signature, i.e., step 407; the data receiver sets a subsequent cache strategy based on the embedded information obtained in step 407, i.e., step 408; the data receiver uses the public key of the data sender to verify the data signature, i.e., step 409; if 409 verification passes, the original data is obtained, i.e., step 410.

[0109] The covert communication path construction method based on the distributed model optimization path of this embodiment integrates technical means such as public-private key pair generation, digital signatures, data encryption, dynamic obfuscation, and protocol field parsing. It aims to solve the problem of data privacy protection in distributed training scenarios and ensure the security, confidentiality, and integrity of data transmission. Through encryption and signature technology, the confidentiality and authenticity of the data are guaranteed; dynamic obfuscation and model optimization path embedding effectively improve the concealment of communication; data segmentation and splicing strategies ensure transmission efficiency; and flexible caching strategies and data integrity verification enhance the adaptability and traceability of the system. This solution achieves privacy protection while ensuring computational efficiency, providing an efficient, secure, and reliable covert communication solution.

[0110] Based on the same general inventive concept, the present invention also protects a covert communication path construction system based on a distributed model optimization path. The covert communication path construction system based on a distributed model optimization path provided by the present invention is described below. The covert communication path construction system based on a distributed model optimization path described below and the covert communication path construction method based on a distributed model optimization path described above can be referenced to each other.

[0111] Figure 4 FIG is a structural diagram of a covert communication path construction system based on distributed model optimization path provided by an embodiment of the present invention. Figure 4 As shown, the covert communication path construction system based on distributed model optimization path of this embodiment includes a data sender 41 and a data receiver 42.

[0112] In a specific implementation process, the data sender 41 digitally signs the original data to generate a data signature, and constructs embedded information based on the original data and the embedding strategy;

[0113] The data sender 41 encrypts the original data, the data signature, and the embedded information to generate data ciphertext, fills the data with a payload according to the channel capacity, divides the data into data blocks, and constructs the protocol fields corresponding to the data blocks;

[0114] The data sender 41 signs the model parameters without embedded information to obtain a parameter signature, and then concatenates the parameter signature, the data block, and the protocol field into a data packet, embeds the pre-built model optimization path, and sends it to the data receiver 42;

[0115] The data receiver 42 extracts the data packet from the model optimization path, verifies the parameter signature to confirm the integrity, and then aggregates the data. After determining that there is a payload in the protocol field, the data block is spliced ​​to obtain the data ciphertext.

[0116] The data receiver 42 decrypts the data ciphertext, decomposes it into the data signature, the original data and the embedded information, verifies the data signature, obtains the original data, and determines the embedding strategy according to the embedded information.

[0117] It should be noted that the relevant information that may be involved in the various embodiments of the present invention are all strictly in accordance with the requirements of laws and regulations, follow the principles of legality, legitimacy and necessity, and are based on reasonable purposes of business scenarios to process information that users actively provide during the use of products / services or generated due to the use of products / services, as well as information obtained with user authorization.

[0118] The relevant information processed by this invention will vary depending on the specific product / service scenario and is subject to the specific scenario in which the user uses the product / service. This may involve the user's account information, device information, or other relevant information. This invention will treat the relevant information and its processing with a high degree of diligence.

[0119] The present invention attaches great importance to the security of relevant information and has adopted reasonable and feasible security protection measures that comply with industry standards to protect relevant information and prevent unauthorized access, public disclosure, use, modification, damage or loss of relevant information.

[0120] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0121] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for constructing a covert communication path based on a distributed model optimization path, characterized in that: include: The data sender digitally signs the original data to generate a data signature, and constructs embedded information based on the original data and the embedding strategy; The data sender encrypts the original data, the data signature, and the embedded information to generate data ciphertext, fills the data with a payload according to the channel capacity, divides the data into data blocks, and constructs the protocol fields corresponding to the data blocks; The data sender signs the model parameters without embedded information to obtain a parameter signature, and concatenates the parameter signature, the data block, and the protocol field into a data packet, embeds the pre-built model optimization path, and sends the data packet to the data receiver; The data receiver extracts the data packet from the model optimization path, verifies the parameter signature to confirm the integrity, and then aggregates the data. After determining that there is a payload in the protocol field, the data block is spliced ​​to obtain the data ciphertext. The data receiver decrypts the data ciphertext, decomposes it into the data signature, the original data and the embedded information, verifies the data signature, obtains the original data, and determines the embedding strategy according to the embedded information.

2. The method for constructing a covert communication path based on a distributed model optimization path according to claim 1, characterized in that: Also includes: If the throughput corresponding to the real-time communication demand is greater than or equal to the preset throughput, both the embedding strategy based on the gradient three-dimensional direction and the embedding strategy based on the gradient amplitude are set as the embedding strategy; If the throughput corresponding to the real-time communication demand is less than the preset throughput, the embedding strategy based on the gradient three-dimensional direction or the embedding strategy based on the gradient amplitude is set as the embedding strategy.

3. The method for constructing a covert communication path based on a distributed model optimization path according to claim 2, characterized in that: The embedding strategy based on the gradient three-dimensional direction is: In the first stage, the trained model parameters are recorded, and an optimization trajectory that can map hidden information is constructed in a three-dimensional coordinate system as the model optimization path; In the second stage, the direction of the gradient corresponding to the model parameters is controlled so that the model optimization path forms a trajectory with set geometric features in the parameter space, thereby embedding the data packet into the model optimization path.

4. The method for constructing a covert communication path based on a distributed model optimization path according to claim 2, characterized in that: The embedding strategy based on gradient magnitude is: In the first stage, the first benchmark linear equation is constructed based on the current model parameters and gradient modulus; In the second stage, a second benchmark linear equation is constructed based on the next-order model parameters and the gradient modulus; Obtaining an initial intersection point of the first reference linear equation and the second reference linear equation; if the position of the initial intersection point meets a preset requirement, using the initial intersection point as the embedding position of the data packet; if the position of the intersection point does not meet the preset requirement, dynamically adjusting the gradient modulus to correct the equation coefficients; and adjusting the position of the intersection point to meet the preset requirement, using the adjusted intersection point as the embedding position of the data packet; The preset requirement condition includes that the intersection point is located in the second quadrant or the third quadrant of the rectangular coordinate system.

5. The method for constructing a covert communication path based on a distributed model optimization path according to claim 1, characterized in that: Before the parameter signature, the data block, and the protocol field are concatenated into a data packet and embedded into a pre-built model optimization path, the method further includes: The data packets are scrambled by using dynamic obfuscation technology to generate data packets in an obfuscated sending queue.

6. The method for constructing a covert communication path based on a distributed model optimization path according to claim 1, characterized in that: Filled with payload according to channel capacity, including: The data length of the data ciphertext is extended to an integer multiple of the effective load of the channel capacity.

7. The method for constructing a covert communication path based on a distributed model optimization path according to claim 1, characterized in that: The protocol field includes a first field and a second field; The first field identifies the load status by the parity of the bit "1", wherein an odd number indicates a valid load and an even number indicates no load; The second field is used to record the sequence number of the data block.

8. The method for constructing a covert communication path based on a distributed model optimization path according to claim 1, characterized in that: The data signature is generated using the ECDSA algorithm and is calculated by the data sender's private key on the original data hash value; The data ciphertext encryption adopts the ECC algorithm and uses the public key of the data receiver to perform the encryption operation.

9. The method for constructing a covert communication path based on a distributed model optimization path according to claim 1, characterized in that: The effective load of the channel capacity is jointly determined by the model architecture and the parameter quantity; the effective load is smaller than the model parameter quantity and includes the total transmission volume of the original data, data signature and protocol information.

10. A covert communication path construction system based on distributed model optimization path, characterized in that: Including data senders and data receivers; The data sender digitally signs the original data to generate a data signature, and constructs embedded information based on the original data and the embedding strategy; The data sender encrypts the original data, the data signature, and the embedded information to generate data ciphertext, fills the data with a payload according to the channel capacity, divides the data into data blocks, and constructs the protocol fields corresponding to the data blocks; The data sender signs the model parameters without embedded information to obtain a parameter signature, and concatenates the parameter signature, the data block, and the protocol field into a data packet, embeds the pre-built model optimization path, and sends the data packet to the data receiver; The data receiver extracts the data packet from the model optimization path, verifies the parameter signature to confirm the integrity, and then aggregates the data. After determining that there is a payload in the protocol field, the data block is spliced ​​to obtain the data ciphertext. The data receiver decrypts the data ciphertext, decomposes it into the data signature, the original data and the embedded information, verifies the data signature, obtains the original data, and determines the embedding strategy according to the embedded information.

Images