An adaptive and privacy-preserving source-channel joint coding online segmentation learning method

Abstract

The application provides a self-adaptive and privacy-protected source channel joint coding online segmentation learning method, comprising the following steps: based on a preset codec network architecture, a segmentation point set containing multiple candidate segmentation points is constructed, each segmentation point divides the encoder into a sending end and a receiving end; based on a context multi-arm tiger machine algorithm, the optimal segmentation point at the current moment is selected from the segmentation point set according to the obtained communication system state information at the current moment; a segmentation reasoning process is performed according to the optimal segmentation point, the sending end obtains intermediate features by processing the input data through the encoder network layer before the optimal segmentation point, and transmits the intermediate features to the receiving end, the receiving end decodes the intermediate features to reconstruct output data; in response to detecting that the data source distribution drifts or the channel condition change amount exceeds a preset threshold, online adaptive segmentation training is triggered. The application realizes the triple balance of reconstruction accuracy, communication overhead and privacy protection.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to an online segmentation learning method for joint source-channel coding with adaptive and privacy-preserving features. Background Technology

[0002] In recent years, with the large-scale access of IoT devices, the surge in demand for high-definition video transmission, and the rapid popularization of edge intelligence applications, traditional communication systems based on separate source-channel coding have struggled to meet increasingly stringent communication requirements due to the significant cliff effect (i.e., a sharp degradation in transmission quality when channel conditions deteriorate). Against this backdrop, deep learning-based Joint Source-Channel Coding (JSCC) technology has emerged and developed rapidly. By training an end-to-end codec network, it jointly optimizes source coding and channel coding, exhibiting superior transmission robustness compared to traditional separate schemes under harsh channel conditions such as low signal-to-noise ratios. It has achieved significant performance improvements under specific datasets and channel conditions, opening a new paradigm for semantic communication. Furthermore, research has begun exploring the online update capabilities of JSCC models to improve environmental adaptability, such as online learning nonlinear transformation source-channel coding schemes, attempting to address data distribution drift and channel variations through continuous parameter fine-tuning.

[0003] However, the existing JSCC solutions mentioned above still have three significant shortcomings in actual deployment: (1) Existing solutions all adopt a fixed end-to-end encoding and decoding architecture. Once the encoder network structure is trained, it is fixed to a single depth. It is impossible to flexibly adjust the encoding computation amount according to the actual computing power of the transmitting device, or to dynamically adapt the encoding depth and compression rate according to the real-time channel status. This makes it difficult to deploy on terminal devices with limited computing power, and faces the problem of efficiency and performance imbalance when the channel conditions change.

[0004] (2) In the transmission process, the intermediate feature map dimension and data distribution of the encoder output in the fixed architecture JSCC model have relatively stable statistical regularity. Eavesdroppers can establish semantic mapping relationship by continuously intercepting the transmission signal and implementing label inference attack, thereby stealing the privacy of the original data. Existing solutions lack an endogenous privacy protection mechanism designed at the architecture level.

[0005] (3) Existing JSCC solutions mostly rely on offline pre-training, and the model parameters are relatively fixed after deployment. Although some works have introduced online fine-tuning, they have failed to incorporate data flow bandwidth overhead, reconstruction distortion and model update flow bandwidth overhead into a unified optimization framework. This makes it difficult to effectively balance the relationship between transmission quality, communication efficiency and model update cost when introducing online adaptive mechanisms, which limits its real-time adjustment capability and deployment flexibility in real dynamic environments. Summary of the Invention

[0006] In view of this, the present invention provides an online segmentation learning method for joint source-channel coding with adaptive and privacy-preserving features. Through dynamic segmentation point selection, online overfitting training and segmentation learning architecture, a triple balance between reconstruction accuracy, communication overhead and privacy protection is achieved.

[0007] This invention provides an online segmentation learning method for adaptive and privacy-preserving source-channel joint coding, comprising: constructing a segmentation point set containing multiple candidate segmentation points based on a preset encoder-decoder network architecture, wherein each segmentation point divides the encoder into a transmitter and a receiver; selecting the optimal segmentation point at the current moment online from the segmentation point set based on a context-based multi-armed slot algorithm, according to the acquired communication system state information at the current moment, wherein the communication system state information includes at least one of the current channel signal-to-noise ratio, recent transmission reconstruction quality, reconstruction quality change rate, and current environment duration; performing a segmentation inference process based on the optimal segmentation point, wherein the transmitter processes the input data through the encoder network layer before the optimal segmentation point to obtain intermediate features and transmits them to the receiver, and the receiver decodes the intermediate features to reconstruct the output data; and triggering online adaptive segmentation training in response to detecting a drift in the data source distribution or a channel condition change exceeding a preset threshold. Optionally, the context-based multi-armed slot machine algorithm selects the optimal split point for the current moment online from the set of split points, including: modeling each candidate split point as an arm, with each arm maintaining historical observation data for online decision-making; at each time step, obtaining a normalized context vector, calculating the upper confidence boundary score for each candidate split point, the upper confidence boundary score being determined based on the sum of the historical reward estimate and the exploration reward; selecting the candidate split point with the largest upper confidence boundary score as the optimal split point for the current moment; and after executing the selected split point, updating the historical observation data corresponding to the split point based on the actual reward obtained.

[0008] Optionally, the actual reward obtained is jointly determined by the positive term of reconstruction quality, the negative term of data flow bandwidth cost, and the negative term of model parameter update cost, wherein the negative term of data flow bandwidth cost and the negative term of model parameter update cost are weighed by their respective weight coefficients.

[0009] Optionally, the selection of the optimal segmentation point is also based on the end-to-end total delay minimization criterion, wherein the end-to-end total delay includes three parts: the encoding calculation delay at the sending end, the intermediate feature transmission delay, and the decoding calculation delay at the receiving end.

[0010] Optionally, the encoding calculation delay at the transmitting end is determined based on the encoding calculation complexity and the processing capability of the transmitting end; the intermediate feature transmission delay is determined based on the amount of intermediate feature data and the achievable rate of the channel; and the decoding calculation delay at the receiving end is determined based on the decoding calculation complexity and the processing capability of the receiving end.

[0011] Optionally, the online adaptive segmentation training includes: freezing all encoder network layers after the optimal segmentation point and all decoder parameters at the receiver, and only performing overfit training on the network parameters of the encoder from the input layer to the optimal segmentation point.

[0012] Optionally, the online adaptive segmentation training is a joint adaptive segmentation training between the transmitter and receiver: the transmitter sends a training request to the receiver, the training request including current channel state information and a new data domain signature, and the receiver confirms participation in the training and the two parties handshake to synchronize; the transmitter trains the network parameters of the encoder from the input layer to the optimal segmentation point, and the receiver trains the network parameters of the decoder and the network parameters of the encoder from the optimal segmentation point to the output layer; in each training iteration, the receiver backfeeds the gradient of the intermediate features to the transmitter, so that the transmitter can complete the gradient calculation and update of the encoder parameters.

[0013] In another aspect, the present invention provides an online segmentation learning system for adaptive and privacy-preserving source-channel joint coding, comprising: a construction module for constructing a segmentation point set containing multiple candidate segmentation points based on a preset encoder-decoder network architecture, wherein each segmentation point divides the encoder into a transmitter and a receiver; a selection module for selecting the optimal segmentation point at the current moment from the segmentation point set online based on a context-based multi-armed slot algorithm and according to the communication system state information obtained at the current moment, wherein the communication system state information includes at least one of the current channel signal-to-noise ratio, recent transmission reconstruction quality, reconstruction quality change rate, and current environment duration; a reconstruction module for performing a segmentation inference process based on the optimal segmentation point, wherein the transmitter processes the input data through the encoder network layer before the optimal segmentation point to obtain intermediate features and transmits them to the receiver, and the receiver decodes the intermediate features to reconstruct the output data; and a training trigger module for triggering online adaptive segmentation training in response to detecting a drift in the data source distribution or a channel condition change exceeding a preset threshold.

[0014] In another aspect, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the steps of the method as described in any of the preceding claims.

[0015] In another aspect, the present invention provides a computer storage medium storing a computer program that, when executed by a processor, implements the steps of the method as described in any of the preceding claims.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) Achieving dynamic adaptation of the JSCC encoding / decoding architecture, effectively enhancing reconstruction quality. This invention utilizes an online segmentation point selection strategy based on the LinUCB algorithm, enabling the system to autonomously switch the optimal coding depth according to real-time channel conditions and data source characteristics, without the need for pre-setting fixed segmentation points or switching schemes. When channel quality improves, a deeper segmentation point is automatically adopted to reduce communication bandwidth overhead; when the channel deteriorates, a shallower segmentation point is automatically switched to enhance error correction redundancy, achieving a dynamic optimal balance between coding efficiency and transmission robustness. Experiments show that in highly fluctuating environments, the reconstruction effect of this invention is significantly higher than that of traditional fixed architecture schemes.

[0017] (2) Enhanced privacy protection through segmentation learning architecture. The original data of this invention is always retained at the sending end, with only intermediate features transmitted through the channel, thus limiting information exposure at the architectural level. Combined with dynamic segmentation point switching and online parameter updates, the feature distribution is continuously perturbed, making it difficult for attackers to establish stable semantic mapping relationships. LIA experiments show that this invention can reduce the attacker's accuracy from approximately 43-49% in traditional schemes to approximately 11-13%, demonstrating significant privacy protection.

[0018] (3) Highly scalable and adaptable to various JSCC architectures. The online segmentation learning framework of this invention is model-independent and has been validated on DeepJSCC, ADJSCC based on CNN, and SwinJSCC based on Transformer, all of which have achieved consistent performance gains. The framework can also be extended to other deep learning encoding and decoding models and different transmission scenarios (such as video transmission, multimodal semantic communication, etc.). Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. By reading the detailed description of the embodiments below, the advantages and benefits of the solutions will become clear to those skilled in the art. The accompanying drawings are only for illustrating preferred embodiments and are not intended to limit the present invention. In the accompanying drawings: Figure 1 This is a flowchart of the steps of the method of the present invention.

[0020] Figure 2 This diagram compares the system architectures of traditional communication systems with those of the online segmentation learning framework proposed in this invention.

[0021] Figure 3 This is a schematic diagram of the 7-layer DeepJSCC architecture and segmentation learning.

[0022] Figure 4 This is a comparison chart of PSNR and communication overhead for different schemes under dynamic environments.

[0023] Figure 5 This is a comparison chart of the reconstruction performance of SwinJSCC on the DIV2K dataset with SNR=1dB. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and thoroughly 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 should fall within the protection scope of the present invention.

[0025] See Figure 1 This invention provides an online segmentation learning method for joint source-channel coding with adaptive and privacy-preserving features, comprising: S1. Based on a preset codec network architecture, construct a segmentation point set containing multiple candidate segmentation points, whereby each segmentation point divides the encoder into a transmitter and a receiver. S2. Based on the context-based multi-armed slot machine algorithm, the optimal segmentation point at the current moment is selected online from the set of segmentation points according to the communication system state information at the current moment. The communication system state information includes at least one of the following: current channel signal-to-noise ratio, recent transmission reconstruction quality, reconstruction quality change rate, and current environment duration. S3. Perform a segmentation inference process based on the optimal segmentation point. The sending end processes the input data through the encoder network layer before the optimal segmentation point to obtain intermediate features and transmits them to the receiving end. The receiving end decodes the intermediate features to reconstruct the output data. S4. In response to the detection of a drift in the data source distribution or a change in channel conditions exceeding a preset threshold, online adaptive segmentation training is triggered.

[0026] Optionally, the context-based multi-armed slot machine algorithm selects the optimal split point for the current moment online from the set of split points, including: modeling each candidate split point as an arm, with each arm maintaining historical observation data for online decision-making; at each time step, obtaining a normalized context vector, calculating the upper confidence boundary score for each candidate split point, the upper confidence boundary score being determined based on the sum of the historical reward estimate and the exploration reward; selecting the candidate split point with the largest upper confidence boundary score as the optimal split point for the current moment; and after executing the selected split point, updating the historical observation data corresponding to the split point based on the actual reward obtained.

[0027] Optionally, the actual reward obtained is jointly determined by the positive term of reconstruction quality, the negative term of data flow bandwidth cost, and the negative term of model parameter update cost, wherein the negative term of data flow bandwidth cost and the negative term of model parameter update cost are weighed by their respective weight coefficients.

[0028] Optionally, the selection of the optimal segmentation point is also based on the end-to-end total delay minimization criterion, wherein the end-to-end total delay includes three parts: the encoding calculation delay at the sending end, the intermediate feature transmission delay, and the decoding calculation delay at the receiving end.

[0029] Optionally, the encoding calculation delay at the transmitting end is determined based on the encoding calculation complexity and the processing capability of the transmitting end; the intermediate feature transmission delay is determined based on the amount of intermediate feature data and the achievable rate of the channel; and the decoding calculation delay at the receiving end is determined based on the decoding calculation complexity and the processing capability of the receiving end.

[0030] Optionally, the online adaptive segmentation training includes: freezing all encoder network layers after the optimal segmentation point and all decoder parameters at the receiver, and only performing overfit training on the network parameters of the encoder from the input layer to the optimal segmentation point.

[0031] Optionally, the online adaptive segmentation training is a joint adaptive segmentation training between the transmitter and receiver: the transmitter sends a training request to the receiver, the training request including current channel state information and a new data domain signature, and the receiver confirms participation in the training and the two parties handshake to synchronize; the transmitter trains the network parameters of the encoder from the input layer to the optimal segmentation point, and the receiver trains the network parameters of the decoder and the network parameters of the encoder from the optimal segmentation point to the output layer; in each training iteration, the receiver backfeeds the gradient of the intermediate features to the transmitter, so that the transmitter can complete the gradient calculation and update of the encoder parameters.

[0032] This invention also provides an online segmentation learning system for adaptive and privacy-preserving joint source-channel coding, comprising: The construction module is used to construct a set of segmentation points containing multiple candidate segmentation points based on a preset codec network architecture, where each segmentation point divides the encoder into a transmitter and a receiver. The selection module is used to select the optimal segmentation point at the current moment from the set of segmentation points online based on the context multi-armed slot machine algorithm and the communication system state information at the current moment. The communication system state information includes at least one of the following: current channel signal-to-noise ratio, recent transmission reconstruction quality, reconstruction quality change rate, and current environment duration. The reconstruction module is used to perform a segmentation inference process based on the optimal segmentation point. The sending end processes the input data through the encoder network layer before the optimal segmentation point to obtain intermediate features and transmits them to the receiving end. The receiving end decodes the intermediate features to reconstruct the output data. The training trigger module is used to trigger online adaptive segmentation training in response to the detection of data source distribution drift or channel condition changes exceeding a preset threshold.

[0033] Specifically, the solution of the present invention is further described with reference to the following examples: See Figure 2 (a) shows the traditional communication system architecture, and (b) shows the online segmentation learning framework proposed in this invention. The online segmentation learning framework of this invention is as follows: System Overall Architecture The core idea of ​​this invention is to split the JSCC codec into a transmitter (Tx) encoder part and a receiver (Rx) decoder part through split learning, and dynamically select the optimal split point and adaptively update the model parameters through an online learning mechanism.

[0034] Unlike traditional end-to-end architectures, this invention, in the encoder's... The segmentation layer (split point) cuts off the network. The network before the split point is deployed at the transmitting end, and the network after the split point is deployed at the receiving end. The output features of the segmentation layer are directly transmitted to the receiving end via the wireless channel, and the receiving end continues to perform decoding operations on the remaining network.

[0035] Unified optimization objective function Based on the theoretical foundation of variational autoencoders (VAEs), this invention models the optimization problem of JSCC online segmentation learning as a three-objective weighted function:

[0036] Among them, the data stream bandwidth cost R reflects the channel resources required for transmitting intermediate features of the encoder output; reconstruction distortion D measures the original data. With reconstruction data Differences between them (such as mean squared error (MSE) or perceptual similarity (LPIPS); model streaming bandwidth cost M is for online updating of model parameters. The objective function establishes that model update overhead is not merely an additional burden, but rather an investment in improving overall rate-distortion efficiency.

[0037] Option 1: Partitioned inference aimed at minimizing end-to-end latency This scheme is suitable for scenarios where the computing power of the sending and receiving ends is asymmetrical, and its goal is to minimize the total end-to-end latency while ensuring transmission quality.

[0038] For each candidate split point The end-to-end total latency is modeled as the sum of three parts:

[0039] in: : The sending end from input to The encoding calculation delay, To calculate the computational complexity, For the sending end's processing capacity; : Transmission delay of intermediate data in the segmentation layer For intermediate feature data volume, The achievable rate of the channel; : The receiving end from Decoding calculation delay to output.

[0040] By solving Selecting the optimal split point. This process can be triggered periodically as the channel or device state changes, achieving dynamic optimization.

[0041] Option 2: Adaptive segmentation training of the transmitter encoder When the transmitter detects a shift in the data source distribution (such as a new image type or sensor) or a significant change in channel conditions (bandwidth, signal-to-noise ratio), it triggers the adaptive segmentation training process.

[0042] The selection criteria for split points introduce a performance-latency tradeoff:

[0043] in, To adapt the encoder to the split point Expected performance improvement and The weighting factor balances the importance of latency and performance.

[0044] Training strategy: Select the optimal split point After that, only the encoder from the input layer to... Overfitting training was performed on the parameters between them, and then the data was frozen. All subsequent network layers:

[0045] Option 3: Joint adaptive segmentation training between the transmitter and receiver This scheme upgrades the centralized decision-making at the transmitter to collaborative optimization between the transmitter and receiver. After detecting a domain offset, the transmitter sends a training request (containing the current channel conditions and the new data domain signature) to the receiver. After the receiver confirms its participation, both parties collaboratively solve for the optimal split point.

[0046] During joint training, the end-to-end task loss is:

[0047] During training, the receiver needs to send back the gradient signal (or loss value) of intermediate features to the transmitter so that the transmitter can complete backpropagation. This bidirectional gradient exchange is the main difference between Scheme 3 and Scheme 2, and its additional communication overhead is exchanged for better end-to-end reconstruction quality.

[0048] Dynamic split point selection based on LinUCB To efficiently select the optimal split point under time-varying channels and data distribution without traversing all candidates, this invention models online split point selection as a contextual multi-armed gambler (MAB) problem.

[0049] System state definition: at time step Construct normalized context vectors It includes: current channel SNR, recent reconstruction PSNR, PSNR change rate, and duration in the current environment.

[0050] Instant reward definition:

[0051] in, Candidate split points The quality of reconstruction and These represent the normalized data flow bandwidth and the model parameter update amount, respectively.

[0052] Selection strategy: Using the LinUCB algorithm, at each time step, select the split point that maximizes the upper confidence boundary.

[0053] in For ridge regression parameter estimation based on historical observations, For the context covariance matrix, To explore the confidence boundary hyperparameters, this algorithm strikes a balance between utilizing historical best split points and exploring potentially better split points.

[0054] Example See Figure 3 This embodiment is based on an improved 7-layer DeepJSCC encoding / decoding architecture. Scheme 3 only fine-tunes the parameters before the sending end segmentation point, while Scheme 4 fine-tunes the parameters at both the sending and receiving ends.

[0055] This embodiment uses the CIFAR-10 and Imagenette datasets, AWGN channels, and PSNR and LPIPS as evaluation metrics. The following are four comparison schemes: Option 1: Original DeepJSCC, fixed architecture; Option 2: Original ADJSCC / SwinJSCC, fixed architecture; Option 3: Adaptive segmentation training of the transmitter encoder (Option 2 of this invention); Option 4: Joint adaptive segmentation training between the transmitter and receiver (Option 3 of this invention).

[0056] Implementation methods for schemes three and four: The implementation process of Scheme 3 (adaptive segmentation training of the transmitter encoder) is as follows: The transmitter monitors data source distribution drift and channel SNR changes at each time step, triggering an adaptive process when the deviation exceeds a threshold. First, the LinUCB algorithm is used to select the optimal segmentation point online from seven candidate segmentation points. Then, freeze. All subsequent encoder layers and all decoder parameters at the receiver are only for encoder layers 1 to 2. Online overfit training is performed on the layers: intermediate features are obtained through forward propagation at the sending end. After transmission through the channel, the receiving end reconstructs the data using a frozen decoder. After calculating the gradient of the reconstructed distortion, it is fed back from the receiver, updating only the trainable parameters at the transmitter. The optimization objective is... After training converges, segmentation inference is performed with the updated encoder until the next environmental change triggers a new round of adaptation.

[0057] Option 4 (Joint Adaptive Segmentation Training at Both Transmitter and Receiver) incorporates the receiver into the training process, building upon Option 3. Upon detecting an environmental change, the transmitter first sends a training request (including the current CSI and the new data domain signature) to the receiver. After confirmation by the receiver, both parties synchronize via a handshake. LinUCB jointly selects the optimal segmentation point. Afterwards, the transmitter trains encoder 1 to... Layer, receiver training decoder To the output layer. In each iteration, the receiver not only calculates its own parameter gradients and updates them... The intermediate feature gradients also need to be sent back to the sender to update the data. Scheme 4 has higher communication overhead (bidirectional gradient flow) than Scheme 3, but the reconstruction quality is better.

[0058] LinUCB Dynamic Split Point Selection This invention models the split point selection as a contextual multi-armed slot machine problem: 7 candidate split points are "arms", one is selected at each time step and the reward is observed, and the optimal strategy is learned online through the LinUCB algorithm.

[0059] Context state: Constructing a normalized feature vector at each step It includes the current channel SNR, the recently reconstructed PSNR, the PSNR change rate, the duration of the environment, and the current segmentation point.

[0060] Reward function: That is, the reconstruction quality minus the penalties for data stream bandwidth and model parameter update amount. , Control and trade-offs.

[0061] Selection Strategy: Each Arm Maintaining the covariance matrix and reward cumulative vector Ridge regression estimation parameters Calculate the UCB score at each step. The first factor is the historical best reward estimate, and the second factor is the bonus for arms that haven't tried enough. The split point with the highest score is selected. Updated with the actual reward increment after execution. and .

[0062] Figure 4 Experimental results validate LinUCB's adaptive capabilities: when the channel deteriorates (e.g., from 10dB to 3dB), it automatically switches from Layer 5 to Layer 3 (CBR increases from 1 / 3 to 2 / 3), trading higher bandwidth for error correction redundancy; when the channel recovers, it automatically switches back to Layer 5 to reduce overhead. This switching is learned online through continuous interaction with the environment, rather than based on predefined rules.

[0063] Core experimental results (1) Reconstruction performance and adaptive segmentation in dynamic environments.

[0064] In a dynamic AWGN transmission environment where data and channel conditions change synchronously, the effectiveness of LinUCB-driven online segmentation learning is verified, and the results are as follows: Figure 4As shown, Schemes 3 and 4 consistently outperform Scheme 1 in PSNR across all time periods, with an average improvement of 2-4 dB. The LinUCB-driven segmentation points autonomously switch when the channel changes (as shown by Switch 1-4 in the figure). Shallow segmentation provides stronger error correction redundancy during low SNR periods, while deep segmentation reduces communication overhead during high SNR periods. Scheme 4 achieves the best PSNR reconstruction, but its communication overhead is slightly higher than that of Scheme 3.

[0065] Furthermore, in highly volatile transmission environments, the average PSNR of Scheme 3 and Scheme 4 remains above 22-23dB, while Scheme 1 degrades to about 18dB. The privacy protection effect is shown in the table below. The LIA attack experiment shows that Scheme 3 reduces the attacker's accuracy from 43-49% of the traditional scheme to about 22-28%, and Scheme 4 further reduces it to about 11-13%.

[0066] Table 1 compares the privacy protection performance of different schemes on the CIFAR10 and Imagenette datasets, quantified by the attacker's average accuracy. A lower value is better.

[0067] Table 1

[0068] (2) Verification of generality on the SwinJSCC model.

[0069] The same framework is applied to SwinJSCC, which is based on Transformer. Figure 5 The reconstruction visual effect of the DIV2K dataset with SNR=1dB is shown. It can be seen that Scheme 4 reconstructs clearer texture details under low signal-to-noise ratio conditions, and the subjective quality is significantly better than the pre-trained baseline, which verifies the universality of this method for different JSCC architectures.

[0070] It should be understood that the adaptive JSCC encoding / decoding framework based on online segmentation learning provided by this invention has the following core functions: Function 1: Dynamic Segmentation Point Selection Technology for DeepJSCC Encoders Based on LinUCB. Based on the LinUCB (Contextual Multi-Arm Slot Machine) algorithm, the system automatically selects the optimal encoder-decoder segmentation point online according to multi-dimensional contextual information such as the current channel signal-to-noise ratio, recent reconstruction quality, PSNR change rate, and environmental duration, achieving dynamic adjustment of coding depth and compression ratio. When the channel quality is good, the system automatically switches to a deeper segmentation point to reduce communication overhead; when the channel quality deteriorates, the system switches to a shallower segmentation point to provide stronger error correction redundancy.

[0071] Function 2: Transmitter-side adaptive segmentation training. When data source distribution drift or significant changes in channel conditions are detected, the transmitter performs overfit training on the encoder based on the current data batch, updating only the encoder parameters from the input layer to the segmentation point, and freezing subsequent network layers. This mechanism enables the system to maintain optimal transmission quality continuously in dynamic environments.

[0072] Function 3: Joint Adaptive Segmentation Training Between Transmitter and Receiver. The transmitter and receiver negotiate the optimal segmentation point through a two-way handshake and jointly perform overfitting training on the encoder and decoder to achieve end-to-end global optimization. This method achieves better reconstruction quality compared to adapting only the transmitter.

[0073] In summary, compared with the prior art, the present invention has the following beneficial effects: (1) Achieving dynamic adaptation of the JSCC encoding / decoding architecture, effectively enhancing reconstruction quality. This invention utilizes an online segmentation point selection strategy based on the LinUCB algorithm, enabling the system to autonomously switch the optimal coding depth according to real-time channel conditions and data source characteristics, without the need for pre-setting fixed segmentation points or switching schemes. When channel quality improves, a deeper segmentation point is automatically adopted to reduce communication bandwidth overhead; when the channel deteriorates, a shallower segmentation point is automatically switched to enhance error correction redundancy, achieving a dynamic optimal balance between coding efficiency and transmission robustness. Experiments show that in highly fluctuating environments, the reconstruction effect of this invention is significantly higher than that of traditional fixed architecture schemes.

[0074] (2) Enhanced privacy protection through segmentation learning architecture. The original data of this invention is always retained at the sending end, with only intermediate features transmitted through the channel, thus limiting information exposure at the architectural level. Combined with dynamic segmentation point switching and online parameter updates, the feature distribution is continuously perturbed, making it difficult for attackers to establish stable semantic mapping relationships. LIA experiments show that this invention can reduce the attacker's accuracy from approximately 43-49% in traditional schemes to approximately 11-13%, demonstrating significant privacy protection.

[0075] (3) Highly scalable and adaptable to various JSCC architectures. The online segmentation learning framework of this invention is model-independent and has been validated on DeepJSCC, ADJSCC based on CNN, and SwinJSCC based on Transformer, all of which have achieved consistent performance gains. The framework can also be extended to other deep learning encoding and decoding models and different transmission scenarios (such as video transmission, multimodal semantic communication, etc.).

[0076] Another aspect of the present invention provides an electronic device, which includes a processor, a memory, a communication bus, and a communication interface.

[0077] in: The processor, memory, and communication interface communicate with each other via a communication bus.

[0078] A communication interface is used to communicate with other electronic devices or servers.

[0079] The processor is used to execute programs, specifically the steps of any of the methods described in the above embodiments.

[0080] Specifically, the program may include program code, which includes computer operation instructions.

[0081] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0082] Memory is used to store programs. Memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive.

[0083] Specifically, the program can be used to cause the processor to execute the steps of any of the methods described in the embodiments. The specific implementation of each step in the program can be found in the corresponding descriptions of the steps and units executed by any of the methods described above, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments.

[0084] An exemplary embodiment of this application also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods of various embodiments of this application.

[0085] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0086] Specific embodiments of the present invention have now been described. Other embodiments are within the scope of the appended claims. In some cases, the actions described in the claims can be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result.

[0087] It should be noted that all directional indications (such as up, down, left, right, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship between the components in a certain order (as shown in the figure). If the specific order changes, the directional indication will also change accordingly.

[0088] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0089] It should be noted that although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of the present invention.

[0090] The examples of the embodiments of the present invention are intended to concisely illustrate the technical features of the embodiments of the present invention, so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to be an improper limitation of the embodiments of the present invention.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive and privacy-preserving source-channel joint coding online segmentation learning method, characterized in that, include: Based on a preset codec network architecture, a set of segmentation points containing multiple candidate segmentation points is constructed, and each segmentation point divides the encoder into a transmitter and a receiver. Based on the contextual multi-armed slot machine algorithm, the optimal segmentation point at the current moment is selected online from the set of segmentation points according to the communication system state information at the current moment. The communication system state information includes at least one of the following: current channel signal-to-noise ratio, recent transmission reconstruction quality, reconstruction quality change rate, and current environment duration. The segmentation inference process is performed based on the optimal segmentation point. The sending end processes the input data through the encoder network layer before the optimal segmentation point to obtain intermediate features and transmits them to the receiving end. The receiving end decodes the intermediate features to reconstruct the output data. In response to the detection of data source distribution drift or channel condition changes exceeding a preset threshold, online adaptive segmentation training is triggered.

2. The method according to claim 1, characterized in that, The context-based multi-armed slot machine algorithm selects the optimal split point at the current moment online from the set of split points, including: Each candidate split point is modeled as an arm, and each arm maintains historical observation data for online decision-making. At each time step, a normalized context vector is obtained, and the upper confidence boundary score of each candidate segmentation point is calculated. The upper confidence boundary score is determined based on the sum of the historical reward estimate and the exploration reward. Select the candidate split point with the highest upper confidence boundary score as the optimal split point at the current time. After executing the selected split point, update the historical observation data corresponding to that split point based on the actual reward obtained.

3. The method according to claim 2, characterized in that, The actual reward obtained is determined by the positive reconstruction quality term, the negative data flow bandwidth cost term, and the negative model parameter update cost term, wherein the negative data flow bandwidth cost term and the negative model parameter update cost term are weighed by their respective weight coefficients.

4. The method according to claim 2, characterized in that, The selection of the optimal segmentation point is also based on the end-to-end total delay minimization criterion, which includes three parts: the encoding calculation delay at the sending end, the intermediate feature transmission delay, and the decoding calculation delay at the receiving end.

5. The method according to claim 4, characterized in that, The encoding calculation delay at the transmitting end is determined based on the encoding calculation complexity and the processing capability of the transmitting end. The intermediate feature transmission delay is determined based on the amount of intermediate feature data and the achievable rate of the channel. The decoding calculation delay at the receiving end is determined based on the decoding calculation complexity and the processing capability of the receiving end.

6. The method according to claim 1, characterized in that, The online adaptive segmentation training includes: Freeze all encoder network layers after the optimal segmentation point and all decoder parameters at the receiver, and only perform overfit training on the network parameters of the encoder from the input layer to the optimal segmentation point.

7. The method according to claim 1, characterized in that, The online adaptive segmentation training is a joint adaptive segmentation training between the transmitter and receiver: The sending end sends a training request to the receiving end. The training request includes current channel state information and a new data field signature. After the receiving end confirms participation in the training, the two parties handshake and synchronize. The transmitting end trains the network parameters of the encoder from the input layer to the optimal segmentation point, and the receiving end trains the network parameters of the decoder and the network parameters of the encoder from the optimal segmentation point to the output layer. In each training iteration, the receiving end sends the gradient of the intermediate features back to the sending end, so that the sending end can complete the gradient calculation and update of the encoder parameters.

8. An online segmentation learning system for joint source-channel coding with adaptive and privacy-preserving features, characterized in that, include: The construction module is used to construct a set of segmentation points containing multiple candidate segmentation points based on a preset codec network architecture, where each segmentation point divides the encoder into a transmitter and a receiver. The selection module is used to select the optimal segmentation point at the current moment from the set of segmentation points online based on the context multi-armed slot machine algorithm and the communication system state information at the current moment. The communication system state information includes at least one of the following: current channel signal-to-noise ratio, recent transmission reconstruction quality, reconstruction quality change rate, and current environment duration. The reconstruction module is used to perform a segmentation inference process based on the optimal segmentation point. The sending end processes the input data through the encoder network layer before the optimal segmentation point to obtain intermediate features and transmits them to the receiving end. The receiving end decodes the intermediate features to reconstruct the output data. The training trigger module is used to trigger online adaptive segmentation training in response to the detection of data source distribution drift or channel condition changes exceeding a preset threshold.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 7.

Images