Semantic security communication method for multiple eavesdroppers

Abstract

The invention relates to the technical field of wireless communication, in particular to a multi-eavesdropper-oriented semantic security communication method, which aims at the problem of semantic privacy leakage caused by multi-eavesdropper collaborative attack, and comprises the following steps of: constructing a MobileJSCC semantic wireless communication network architecture, efficiently extracting semantic features by adopting joint source channel coding (JSCC), and establishing a multi-eavesdropper semantic security communication network architecture; designing a multi-eavesdropping confrontation joint loss function; by jointly optimizing the reconstruction loss of the legal end and the interference loss of the eavesdropping end, high-quality image recovery of the legal receiving end is realized, and meanwhile, an eavesdropper is prevented from decoding core semantic content; finally, a large number of experiments are carried out on an additive white Gaussian noise channel and a multi-data set, and the semantic privacy protection effectiveness of the method in a multi-eavesdropping scene is verified. The method is suitable for resource-constrained equipment, has both transmission efficiency and security, and meets the requirements of sensitive information transmission.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communications, and in particular to a semantically secure communication method oriented to multiple eavesdroppers. Background Art

[0002] With the large-scale commercialization of 5G technology and the accelerated advancement of 6G research, the demand for transmission efficiency and intelligence in wireless communication systems is growing exponentially. Traditional communication systems, based on the Shannon information theory framework, focus on accurately transmitting bit-level symbols. However, this mechanism has significant drawbacks: in bandwidth-constrained scenarios, redundant bit-level transmission leads to low spectral efficiency. Against this backdrop, semantic communication has emerged as a next-generation communication paradigm. Its core concept is to achieve efficient transmission by extracting the semantic features of information, allowing the receiver to reconstruct the information based on semantic understanding rather than mechanically reconstructing the bit stream. This technology significantly improves transmission efficiency and is particularly suitable for resource-constrained scenarios such as the Internet of Things and edge computing.

[0003] However, the evolution of semantic communication technology has led to new security threats arising from its unique semantic-level information exposure. Traditional communication system security mechanisms focus on protecting the physical layer bitstream, relying on encryption algorithms and channel coding techniques to ensure the confidentiality of data transmission. However, semantic communication abandons the strict bit-level symbol transmission paradigm and instead achieves efficient communication through semantic feature compression. This expands the attack surface from bit leakage to semantic leakage. Therefore, eavesdroppers no longer need to decipher the entire data stream; they only need to recover key semantic features to reconstruct sensitive information.

[0004] Currently, most semantic communication security research focuses on the scenario of a single eavesdropper. However, in real-world applications, eavesdroppers are often not isolated. Multiple eavesdroppers may collaborate, overcoming the limitations of a single eavesdropper through information sharing and joint processing, and reconstructing a more complete message than could be achieved by individual eavesdroppers. Therefore, ensuring the security of semantic communication in a multi-eavesdropper environment has become a pressing issue. Summary of the Invention

[0005] The purpose of the present invention is to provide a semantically secure communication method for multiple eavesdroppers. Through a lightweight network architecture and a multi-objective joint optimization mechanism, the problem of semantic privacy leakage caused by collaborative attacks by multiple eavesdroppers is solved. While ensuring high-fidelity reconstruction at the legitimate receiving end, the semantic recovery capability of the eavesdropping end is directionally suppressed, and the deployment requirements of resource-constrained devices are adapted.

[0006] To achieve the above object, the present invention provides a semantically secure communication method for multiple eavesdroppers, comprising the following steps:

[0007] Step 1: Build the MobileJSCC semantic wireless communication network architecture, which includes the MobileJSCC encoder, wireless channel, and MobileJSCC decoder;

[0008] Step 2: Use the MobileJSCC encoder to encode the semantic features of the original image to generate symbols containing semantic information;

[0009] Step 3: Transmit semantic information via wireless channel broadcast;

[0010] Step 4: Decode the received signal using the MobileJSCC decoder to restore the original image;

[0011] Step 5: Design a joint loss function for multi-eavesdropping resistance. By jointly optimizing the reconstruction loss of the legitimate receiver and the interference loss of the eavesdropper, high-quality image restoration is achieved at the legitimate receiver while preventing the eavesdropper from decoding the core semantic content.

[0012] Step 6: In the AWGN channel and multi-dataset test, verify the privacy protection effectiveness of the semantic secure communication method for multiple eavesdroppers in multiple eavesdropping scenarios.

[0013] Optionally, the MobileJSCC encoder includes a standard convolutional layer, two depth-wise separable convolutional layers, and two inverted residual depth-wise separable convolutional layers connected in sequence, and PReLU activation function and normalization layer processing are used between each layer to achieve efficient semantic feature extraction in resource-constrained devices.

[0014] Optionally, the wireless channel is an additive white Gaussian noise channel, that is, an AWGN channel.

[0015] Optionally, the MobileJSCC decoder includes five cascaded depth-separable transposed convolutional layers, which restore the image spatial resolution through a step-by-step upsampling operation and ultimately output the reconstructed original image.

[0016] Optionally, the expression of the multi-eavesdropping countermeasure joint loss function L is:

[0017] L=λL legit +(1-λ)L eav

[0018] λ∈[0,1] is the global balance parameter between the legitimate end and the eavesdropping end losses;

[0019] The reconstruction loss function L at the legitimate receiving end legit for:

[0020]

[0021] Eavesdropping interference loss function Leav for:

[0022]

[0023] in:

[0024] α∈[0,1] is the weight parameter of mean square error MSE and structural similarity index measurement SSIM;

[0025] Reconstructing the image for the legitimate side With the original image x i MSE between

[0026] Reconstructing the image for the legitimate side With the original image x i SSIM between;

[0027] Reconstructing images for multiple eavesdroppers and The maximum value MAX between them;

[0028] The eavesdropping end has a completely black image 0 and MSE between

[0029] The eavesdropping end has a completely black image 0 and SSIM between;

[0030] B is the batch processing volume;

[0031] N is the number of eavesdroppers.

[0032] Optionally, the MobileJSCC semantic wireless communication network architecture dynamically selects the loss function according to the privacy protection requirements: in the public broadcast scenario, only the loss function L is used. legit , enabling the multi-eavesdropping adversarial joint loss function L in the privacy protection scenario.

[0033] Optionally, the MobileJSCC encoder and decoder optimize model parameters through end-to-end joint training, and the training objective function is min θ,δ L, where θ is the MobileJSCC encoder parameter and δ is the MobileJSCC decoder parameter.

[0034] This paper provides a semantically secure communication method for multiple eavesdroppers. To address the semantic privacy leakage caused by coordinated attacks by multiple eavesdroppers, the paper builds a MobileJSCC semantic wireless communication network architecture, employs joint source-channel coding (JSCC) to efficiently extract semantic features, and designs a joint loss function to combat multiple eavesdroppers. By jointly optimizing the reconstruction loss of the legitimate end and the interference loss of the eavesdropper, the paper achieves high-quality image recovery at the legitimate receiving end while preventing eavesdroppers from decoding the core semantic content. Finally, extensive experiments conducted on additive white Gaussian noise channels and multiple datasets verify the effectiveness of this method in semantic privacy protection in multiple eavesdropper scenarios. The paper is suitable for resource-constrained devices, achieving both transmission efficiency and security, meeting the requirements for sensitive information transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only 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.

[0036] Figure 1 The present invention is a flowchart of the steps of a semantically secure communication method for multiple eavesdroppers.

[0037] Figure 2 It is a schematic diagram of the semantic communication environment of wireless image transmission in a multi-eavesdropping scenario in a specific embodiment of the present invention.

[0038] Figure 3 This is a schematic diagram of the neural network structure of the MobileJSCC encoder and the MobileJSCC decoder in a semantic secure communication method for multiple eavesdroppers of the present invention.

[0039] Figure 4 3 is a schematic diagram comparing the PSNR performance of a legitimate terminal and an eavesdropping terminal under an AWGN channel according to a specific embodiment of the present invention.

[0040] Figure 5 3 is a schematic diagram comparing the PSNR performance of a legitimate terminal and an eavesdropping terminal under an AWGN channel according to a specific embodiment of the present invention.

[0041] Figure 6 Schematic diagram of image restoration quality of the legitimate end and the eavesdropping end on the CIFAR10 dataset according to a specific embodiment of the present invention.

[0042] Figure 7 The figure is a schematic diagram of image restoration quality of a legitimate end and an eavesdropping end on the Kodak dataset according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0044] See also Figure 1 The present invention provides a semantically secure communication method for multiple eavesdroppers, comprising the following steps:

[0045] Step 1: Build the MobileJSCC semantic wireless communication network architecture, which includes the MobileJSCC encoder, wireless channel, and MobileJSCC decoder;

[0046] Step 2: Use the MobileJSCC encoder to encode the semantic features of the original image to generate symbols containing semantic information;

[0047] Step 3: Transmit semantic information via wireless channel broadcast;

[0048] Step 4: Decode the received signal using the MobileJSCC decoder to restore the original image;

[0049] Step 5: Design a joint loss function for multi-eavesdropping resistance. By jointly optimizing the reconstruction loss of the legitimate receiver and the interference loss of the eavesdropper, high-quality image restoration is achieved at the legitimate receiver while preventing the eavesdropper from decoding the core semantic content.

[0050] Step 6: In the AWGN channel and multi-dataset test, verify the privacy protection effectiveness of the semantic secure communication method for multiple eavesdroppers in multiple eavesdropping scenarios.

[0051] The following is further explained with reference to specific embodiments and execution steps:

[0052] like Figure 2 As shown, in this embodiment, a semantic communication environment for wireless image transmission in a multi-eavesdropping scenario is constructed. The sender extracts the semantic feature y to be transmitted from the original image x:

[0053] y=f(x ; θ) (1)

[0054] Where f represents a function mapping relationship for extracting semantic features y from the original image x, and θ represents the trainable parameters of the MobileJSCC encoder at the sender.

[0055] After transmission through the AWGN channel, the feature map received by the legitimate receiving end Bob (in secure communication, the legitimate receiving end is usually named Bob; the eavesdropping end is named Eve. In this embodiment, two eavesdropping ends are used, so they are named Eve1 and Eve2) for:

[0056]

[0057] where h b is the channel gain of the channel between the sender and the legitimate receiver Bob, n b is additive white Gaussian noise of the same dimension as y, and the symbol ⊙ represents the dot product operation.

[0058] The legitimate receiver Bob uses the MobileJSCC decoder to decode the received image features:

[0059]

[0060] in represents the reconstructed image output by the MobileJSCC decoder, and g represents a semantic feature received from Reconstruct the image δ represents the trainable parameters of the MobileJSCC decoder at the legitimate receiving end Bob.

[0061] Due to the openness of wireless channels, signals are received by both legitimate receivers and potential eavesdroppers simultaneously. When there are multiple eavesdroppers, they may collaborate to eavesdrop or attack. By sharing information and conducting joint processing, multiple eavesdroppers can overcome the limitations of a single eavesdropper and reconstruct a more complete message than a single eavesdropper. This embodiment considers the case where there are two eavesdroppers. The eavesdroppers conspire to obtain the best restored image by monitoring the communication signal. The image they receive is and They are:

[0062]

[0063] in, and is the channel gain of the eavesdropping channel of the two eavesdropping terminals (Eve1 and Eve2), and is the additive white Gaussian noise of the eavesdropping channel of the two eavesdropping terminals (Eve1 and Eve2).

[0064] The eavesdropping terminals (Eve1 and Eve2) decode the received images using the MobileJSCC decoder and obtain the following:

[0065]

[0066] Finally, collude to obtain the best eavesdropping image, that is, to obtain the reconstructed image of the two eavesdropping ends and The maximum value (MAX) between:

[0067]

[0068] like Figure 3 As shown in (a), the MobileJSCC encoder adopts a depthwise separable convolutional model, enabling semantic communication on resource-limited local terminal devices. The core task of the encoder is to convert the input image into a compressed feature map and extract image features through a series of convolutional layers. First, the data entering the MobileJSCC encoder passes through a standard convolutional layer to prevent excessive loss of semantic information; then, a depthwise separable convolutional (DSC) layer is used, which separates spatial and channel operations to improve computational efficiency; then, a depthwise separable convolutional (ResDSC) layer with inverted residual connections is used to further learn deeper and more complex features. Each convolutional layer is followed by a PReLU activation function, and finally passes through a normalization layer to meet the system's average transmission power constraint.

[0069] like Figure 3 As shown in (b), the MobileJSCC decoder uses five depth-separable transposed convolution layers (DSTC) to gradually restore image details and ultimately reconstruct the original image. The entire system uses end-to-end joint optimization to effectively achieve efficient image transmission and restoration.

[0070] The loss function is used to measure the gap between the prediction results of the neural network model and the true value. The traditional mean square error (MSE) loss function is defined as: in, To reconstruct the image, I i is the original image, and B is the batch size. Although MSE can effectively constrain pixel-level errors, it suffers from the defect of missing structural information: it only focuses on pixel differences and ignores high-level semantic features such as image texture and edges.

[0071] This paper addresses the multi-objective requirements of semantic communication systems and proposes a hierarchical loss function design framework that balances system universality, confidentiality, and scenario adaptability. This framework achieves dual optimization of high-fidelity reconstruction on the legitimate end and semantic interference on the eavesdropping end by dynamically combining pixel-level error, structural similarity metrics, and a multi-eavesdropping countermeasure mechanism. The specific design is as follows:

[0072] The structural similarity index (SSIM) is introduced as the perceptual loss to form a hybrid objective function with MSE, and the reconstruction loss function L at the legal receiving end is legit for:

[0073]

[0074] in:

[0075] α∈[0,1] is the weight parameter of MSE and SSIM;

[0076] Reconstructing the image for the legitimate side With the original image x i MSE between

[0077] Reconstructing the image for the legitimate side With the original image x i SSIM between;

[0078] Through the loss function L legit By jointly optimizing MSE and SSIM, the proposed system reduces pixel errors while enhancing the ability to preserve image structural features, significantly improving the visual fidelity of the reconstructed image on the legitimate end.

[0079] Furthermore, for the scenario of coordinated attack by multiple eavesdroppers, this paper proposes an adversarial interference mechanism. By jointly optimizing the legitimate end reconstruction and the eavesdropping end interference, a multi-eavesdropping adversarial joint loss function L is constructed:

[0080] L=λL legit +(1-λ)L eav (10)

[0081] Among them, λ∈[0,1] is the global balance parameter of the loss between the legitimate end and the eavesdropping end; L legit Reconstruct the loss function for the legitimate receiver, L eav is the interference loss function of the eavesdropping end:

[0082]

[0083] Where: α∈[0,1] is the weight parameter of MSE and SSIM;

[0084] Reconstructing images for multiple eavesdroppers and The maximum value between (MAX);

[0085] The eavesdropping end has a completely black image 0 and MSE between

[0086] The eavesdropping end has a completely black image 0 and SSIM between;

[0087] B is the batch processing volume;

[0088] N is the number of eavesdroppers.

[0089] By adjusting the global balance parameter λ, the present invention dynamically balances the reconstruction accuracy of the legitimate end and the interference strength of the eavesdropping end during training.

[0090] To adapt to different application scenarios, the present invention supports dynamic loss function switching:

[0091] Public Broadcast Mode: Only L legit Loss function, giving priority to ensuring transmission efficiency;

[0092] Privacy protection mode: Enable the multi-eavesdropping adversarial joint loss function L and activate the multi-eavesdropping adversarial mechanism.

[0093] This embodiment adopts an end-to-end joint training strategy. The MobileJSCC encoder parameters θ and the MobileJSCC decoder parameters δ are trained by the objective function min θ,δ L is jointly optimized. A gradient reversal layer is used during training to invert the eavesdropping end gradient during back propagation, further enhancing the anti-interference effect.

[0094] In this embodiment, the CIFAR-10 image dataset is used to train and evaluate the neural network. CIFAR-10 is a widely used image classification dataset that contains RGB images in 10 categories, with 5,000 training images and 1,000 test images for each category. 50,000 RGB images with a resolution of 32×32 were selected from the CIFAR10 dataset as training data. These images cover different object categories, such as airplanes, cars, birds, cats, etc., and have a certain degree of diversity and complexity, making them suitable for training models of semantic communication systems. In order to evaluate the performance of the model, this embodiment is evaluated on 10,000 test images of the CIFAR10 dataset, and the eavesdropper's eavesdropping performance and the effectiveness of the proposed defense method are compared. This evaluation helps to verify the effectiveness of the proposed method on real data, especially its robustness in the face of eavesdropper attacks.

[0095] In order to comprehensively evaluate the quality of image restoration, the present invention uses two common performance indicators: peak signal-to-noise ratio (PSNR) and perceptual image block similarity (LPIPS). PSNR is a traditional indicator for measuring image restoration quality. The higher the value, the more similar the restored image is to the original image. In this embodiment, the image quality is evaluated by calculating the PSNR value between the restored image and the original image. A higher PSNR value generally means fewer errors and better detail retention during the image restoration process. LPIPS is an indicator used for image perceptual quality assessment in recent years. It measures the similarity between images by calculating the difference in the depth features of the images. Compared with PSNR, LPIPS can better reflect the difference in image perception in the human eye. A lower LPIPS value indicates that the restored image is more similar to the original image in human eye perception, and vice versa, the difference is larger. In this embodiment, LPIPS is used to evaluate the perceptual quality of the restored image, especially in the complex background and detail performance.

[0096] In the experiment, the learning rate is set to 0.001 and training is performed for 500 rounds. It is assumed that both the communication channel and the eavesdropping channel are AWGN channels, and the default bandwidth compression ratio is set to 0.5.

[0097] like Figure 4 As shown, under the AWGN channel, the loss function L in the public broadcast mode is used legit When , the reconstruction performance of the legitimate receiver Bob reaches the optimal PSNR indicator, but the PSNR value of the eavesdropping end Eve is also high at this time, and shows an upward trend as the signal-to-noise ratio (SNR) increases. This shows that the broadcast mode has the risk of semantic privacy leakage. After adopting the multi-eavesdropping countermeasure joint loss function L, although Bob's PSNR value is slightly lower than the performance in the broadcast mode when SNR>10dB, it has a significant inhibitory effect on the reconstruction quality of the two eavesdropping ends (Eve1 and Eve2). Figure 4 As shown in the figure, the PSNR values ​​of Eve1 and Eve2 never exceed 7dB and do not improve with the increase of SNR, while Bob's PSNR always remains above 27dB. This result verifies the semantic privacy protection capability of the proposed method in the multi-eavesdropping scenario.

[0098] Because the loss function incorporates SSIM constraints, this experiment did not directly use SSIM as the evaluation metric to avoid overlap between the evaluation metric and the training objective, and thus causing discrepancies. LPIPS uses a deep learning model to extract differences in image features, which is more consistent with human visual perception. Lower LPIPS values ​​indicate higher image similarity.

[0099] like Figure 5As shown in the figure, in broadcast mode, Bob's LPIPS value remains below 0.1, while the LPIPS value of Eve, the eavesdropper, gradually decreases with increasing SNR, indicating that the eavesdropper's reconstruction quality improves, which is contrary to the goal of privacy protection. In contrast, after adopting the multi-eavesdropping adversarial joint loss function, Bob's LPIPS value remains stable below 0.1, while Eve1 and Eve2's LPIPS values ​​are significantly higher than Bob's (Eve1 fluctuates between 0.4 and 0.6). This fluctuation is due to the eavesdropper's ability to randomly restore only a portion of pixels, resulting in unstable perceptual similarity.

[0100] like Figure 6 As shown in the figure, on the test images of the CIFAR-10 dataset, the reconstruction results of the multi-eavesdropping adversarial joint loss function are consistent with expectations: the legitimate receiver can completely restore the image, while the strong eavesdropper (Eve1) can only reconstruct some pixels, and the weak eavesdropper (Eve2) can hardly recover any valid information. This shows that this method can effectively prevent semantic privacy leakage for different image contents.

[0101] like Figure 7 As shown in the figure, a model trained on the ImageNet 2012 dataset (batch size 1024, learning rate 0.001, and training for 20 epochs) achieved excellent PSNR, LPIPS, and visual perception quality at the legitimate receiver (PSNR ≥ 27dB, LPIPS < 0.1), while the reconstruction results at the eavesdropper were severely distorted. The Kodak dataset contains 24 high-resolution images whose complex textures and details pose a greater challenge to the model. The experimental results verify the universality of the proposed method across images of varying resolutions.

[0102] The above disclosure is merely one or more preferred embodiments of the present invention, and certainly cannot be used to limit the scope of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made in accordance with the claims of the present invention still fall within the scope of the invention.

Claims

1. A semantically secure communication method for multiple eavesdroppers, characterized in that: The following steps are involved: Step 1: Build the MobileJSCC semantic wireless communication network architecture, which includes the MobileJSCC encoder, wireless channel, and MobileJSCC decoder; Step 2: Use the MobileJSCC encoder to encode the semantic features of the original image to generate symbols containing semantic information; Step 3: Transmit semantic information via wireless channel broadcast; Step 4: Decode the received signal using the MobileJSCC decoder to restore the original image; Step 5: Design a joint loss function for multi-eavesdropping resistance. By jointly optimizing the reconstruction loss of the legitimate receiver and the interference loss of the eavesdropper, high-quality image restoration is achieved at the legitimate receiver while preventing the eavesdropper from decoding the core semantic content. Step 6: In the AWGN channel and multi-dataset test, verify the privacy protection effectiveness of the semantic secure communication method for multiple eavesdroppers in multiple eavesdropping scenarios.

2. The semantically secure communication method for multiple eavesdroppers according to claim 1, wherein: The MobileJSCC encoder includes a standard convolutional layer, two depth-wise separable convolutional layers, and two inverted residual depth-wise separable convolutional layers connected in sequence. PReLU activation function and normalization layer processing are used between each layer to achieve efficient semantic feature extraction in resource-constrained devices.

3. The semantically secure communication method for multiple eavesdroppers according to claim 2, wherein: The wireless channel is an additive white Gaussian noise channel, that is, an AWGN channel.

4. The semantically secure communication method for multiple eavesdroppers according to claim 3, wherein: The MobileJSCC decoder contains five cascaded depth-wise separable transposed convolutional layers, which restores the image spatial resolution through step-by-step upsampling operations and finally outputs the reconstructed original image.

5. The semantically secure communication method for multiple eavesdroppers according to claim 4, wherein: The expression of the multi-eavesdropping countermeasure joint loss function L is: L=λL legit +(1-λ)L eav λ∈[0,1] is the global balance parameter between the legitimate end and the eavesdropping end losses; The reconstruction loss function L at the legitimate receiving end legit for: Eavesdropping interference loss function L eav for: in: α∈[0,1] is the weight parameter of mean square error MSE and structural similarity index measurement SSIM; Reconstructing the image for the legitimate side With the original image x i MSE between Reconstructing the image for the legitimate side With the original image x i SSIM between; Reconstructing images for multiple eavesdroppers and The maximum value MAX between them; The eavesdropping end has a completely black image 0 and MSE between The eavesdropping end has a completely black image 0 and SSIM between; B is the batch processing volume; N is the number of eavesdroppers.

6. The semantically secure communication method for multiple eavesdroppers according to claim 5, characterized in that: The MobileJSCC semantic wireless communication network architecture dynamically selects the loss function according to the privacy protection requirements: in the public broadcast scenario, only the loss function L is used. legit , enabling the multi-eavesdropping adversarial joint loss function L in the privacy protection scenario.

7. The semantically secure communication method for multiple eavesdroppers according to claim 6, wherein: The MobileJSCC encoder and decoder optimize model parameters through end-to-end joint training, and the training objective function is min θ,δ L, where θ is the MobileJSCC encoder parameter and δ is the MobileJSCC decoder parameter.

Images