Semantic image transmission method based on homomorphic encryption and deep joint source channel coding

Abstract

This invention proposes a semantic image transmission method based on homomorphic encryption and deep joint source-channel coding. The method includes: at the transmitting end, extracting the latent semantic representation of the image using a semantic encoder, obtaining integer plaintext through linear scaling and quantization, generating semantic ciphertext using BFV homomorphic encryption, and then transmitting it after Gray-coded QAM modulation; at the receiving end, hard-demodulating the signal, decrypting it with the BFV private key to obtain integer plaintext perturbed by channel error and decryption noise, recovering the continuous semantic representation through inverse quantization, and then reconstructing the output image using a semantic decoder. This invention utilizes the additive noise characteristic of BFV decryption error, allowing decryption noise and channel noise to naturally superimpose. The system reconstruction quality smoothly decreases as channel noise increases, overcoming the "cliff effect" of traditional modular encryption schemes where a slight increase in channel error triggers catastrophic decoding or decryption failure under finite code length or time-varying channel conditions. This achieves a balance between cryptographic security and the graceful degradation characteristics of semantic communication.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication and information security technology, and in particular to a semantic image transmission method based on homomorphic encryption and deep joint source channel coding. Background Technology

[0002] With the rapid development of wireless communication technology, image transmission is increasingly widely used in fields such as autonomous driving, telemedicine, security monitoring, and military reconnaissance. However, the open nature of wireless channels exposes transmitted data to various security threats, including eavesdropping, tampering, and forgery. Traditional communication systems, based on Shannon's separation theorem, design source coding, channel coding, and encryption as cascaded independent modules. While this modular design is easy to implement, it suffers from performance bottlenecks under conditions of limited code length, strict delay constraints, or time-varying channels: a slight increase in channel error can trigger catastrophic decoding or decryption failures, known as the "cliff effect."

[0003] In recent years, Deep Learning-based Joint Source-Channel Coding (Deep JSCC) technology has directly mapped source data to channel symbols through end-to-end learning, enabling a smooth decline in reconstruction quality as channel conditions deteriorate, demonstrating inherent robustness to channel noise. Among these, schemes such as WITT based on the Swin Transformer have achieved significantly better performance than traditional schemes in high-resolution image transmission.

[0004] However, existing Deep JSCC schemes generally lack cryptographic security protection. Since the symbols transmitted by Deep JSCC are directly related to the source image, an eavesdropper, after intercepting these symbols, can reconstruct the original image to some extent, even without additional information. Existing security enhancements are mostly isolated patches targeting the physical layer, lacking cryptographic embedding at the system architecture level. Homomorphic encryption supports direct arithmetic operations on ciphertext, and its decryption error is additive noise, naturally compatible with the smooth degradation characteristics of JSCC. However, the homomorphic encryption pipeline is completely non-differentiable, cutting off the end-to-end gradient flow, causing the encoder to be unable to learn semantic representations robust to encryption-channel composite noise.

[0005] In summary, existing wireless image transmission schemes have the following technical problems in integrating cryptographic security and deep joint source-channel coding: 1) lack of architecture-level security protection, making them vulnerable to eavesdropping and model inversion attacks; 2) the non-differentiable barrier when homomorphic encryption and JSCC are integrated prevents the encoder from participating in end-to-end optimization; 3) the small precision loss introduced by encryption actually reduces reconstruction fidelity under high signal-to-noise ratio, i.e. there is a trade-off between security and fidelity. Summary of the Invention

[0006] This invention proposes a semantic image transmission method based on homomorphic encryption and deep joint source channel coding, aiming to solve the following technical problems: 1) the cliff effect caused by traditional modular encryption schemes in JSCC; 2) the non-differentiability of the homomorphic encryption pipeline hinders the joint optimization of the encoder and decoder; 3) the problem that existing schemes have insufficient robustness at low signal-to-noise ratio and loss of fidelity at high signal-to-noise ratio.

[0007] The technical solution of this invention is implemented as follows: This invention provides a semantic image transmission method based on homomorphic encryption and deep joint source-channel coding, comprising the following steps: S1, the sending end extracts the input image into a semantic latent representation through a semantic encoder, and performs linear scaling and quantization on the semantic latent representation to obtain integer plaintext; S2, the sending end uses the BFV homomorphic encryption scheme to encrypt the integer plaintext and generate semantic ciphertext; S3, the transmitting end converts the semantic ciphertext into a binary sequence, performs Gray-coded QAM modulation, generates baseband symbols, and transmits them through the wireless channel; S4, the receiving end performs hard decision demodulation on the received signal to recover the noisy semantic ciphertext; S5, the receiver uses the BFV private key to decrypt the noisy semantic ciphertext, obtaining the integer plaintext that is jointly disturbed by channel error and BFV decryption noise, and then recovers the continuous semantic representation through inverse quantization; S6, the receiving end reconstructs the continuous semantic representation into an output image through a semantic decoder.

[0008] Specifically, the method also includes an encrypted gradient tunneling mechanism: A first pass-through estimator node is set at the output of the quantizer at the transmitting end, and a second pass-through estimator node is set at the input of the dequantizer at the receiving end. A gradient bypass channel that bypasses the encryption, modulation, channel transmission, demodulation and decryption steps is constructed to backpropagate the error signal at the receiving end to the semantic encoder at the transmitting end, so as to achieve end-to-end joint optimization.

[0009] Furthermore, the method also includes a joint encryption fine-tuning strategy: Using the gradient bypass channel, the error between the reconstructed image and the input image is used as the loss function, and the parameters of the semantic encoder and semantic decoder are used as optimization variables. End-to-end joint fine-tuning of the encoder and decoder is performed by minimizing the loss function; the fine-tuning objective function is expressed as: ; in, and These are the parameters for the encoder and decoder, respectively. For the input image, The channel signal-to-noise ratio (SNR) For loss function, For decoder, This is a semantic representation affected by complex noise. During the joint encryption fine-tuning process, the soft-assignment temperature parameter of the quantizer increases from its initial value to its maximum value according to a step-linear temperature ramping schedule, which is expressed as follows: ; in, This represents the current number of training steps. This represents the total number of steps. The starting temperature, This is the scaling factor. The maximum temperature, This indicates rounding down to the nearest integer.

[0010] Preferably, the method further includes a bit error rate gating inference mechanism: After hard decision demodulation, the receiver measures the bit error rate of the current transmission block in real time and compares the bit error rate with a pre-calibrated threshold value. If the bit error rate is higher than the threshold, the model fine-tuned by joint encryption is selected for reconstruction; If the bit error rate is not higher than the threshold, then the pre-trained model is selected for reconstruction.

[0011] Furthermore, the threshold value is calibrated in the following manner: The average peak signal-to-noise ratio of the pre-trained model and the fine-tuned model, as well as the average bit error rate experienced by the pre-trained model, are recorded at a set of signal-to-noise ratio sampling points to calculate the performance advantage. Find the crossover interval where the performance advantage of the fine-tuned model relative to the pre-trained model turns from positive to negative, and calculate the optimal bit error rate threshold within the crossover interval using linear interpolation.

[0012] Specifically, in step S1, the quantization operation is as follows: Using the global minimum determined during the pre-training phase and maximum value The semantic latent representation is linearly scaled to the [0, 255] interval, and the scaling formula is: ; in, and These are the semantic latent representations before and after scaling, respectively; Then perform the rounding operation: At the first straight-through estimation node of the encrypted gradient tunneling mechanism, hard rounding is performed during forward propagation. , This represents the semantic latent representation after hard rounding, i.e., the plaintext integer; Backpropagation employs distance-based soft-quantization surrogate gradients, where the soft-quantization weights are calculated using the following formula: , ; in, For temperature parameters, Scaling value With the center of mass The square of the Euclidean distance; Softening value The calculation formula is: ; The gradient of the soft path is backpropagated to the encoder as a gradient estimate with hard rounding.

[0013] Specifically, in step S2, the ciphertext generation method for BFV homomorphic encryption is as follows: For each plaintext integer Encoded as a polynomial The constant term is The remaining coefficients are zero; the sampling noise vector is , , Calculate the ciphertext pair: ; ; in, and For encrypted pairs, For the common matrix, For private key, To and The corresponding public key component, The error distribution is discrete Gaussian. For the ciphertext modulus, and Let be the dimension of the noise vector. The length of the plaintext vector.

[0014] Specifically, in step S4, the hard decision demodulation rule is as follows: ; in, The QAM constellation points are estimated after hard decision demodulation. For the received first A symbol, For the baseband symbol to be transmitted, For the 64-QAM constellation point set; The receiving end will Mapping back to Gray code labels, removing padding bits, and then reassembling yields semantically ciphertext containing noise.

[0015] Specifically, in step S5, the inverse quantization employs a software dequantization operation: For the decrypted integer plaintext Calculate its relationship with each centroid. square Euclidean distance : ; The soft-assignment weights are obtained by temperature-weighted softmax. : , ; in, For temperature parameters; Soft dequantization outputs continuous values for: ; During backpropagation, the gradient is propagated back using the following formula: ; gradient It is directly routed to the first direct-access estimator node via the identity mapping link.

[0016] Specifically, the semantic encoder adopts a hierarchical Swing Transformer structure, which includes four stages. Each stage performs spatial downsampling and semantic feature transformation in sequence. After the last stage, a learnable 1×1 convolutional layer is connected to compress the number of channels of the feature map to the dimension corresponding to the target semantic channel bandwidth ratio. The semantic decoder adopts a Swing Transformer structure symmetrical to the semantic encoder, which gradually restores the feature map to the original resolution through four upsampling stages.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention utilizes the characteristic that the BFV decryption error is additive noise, so that the decryption noise and the channel noise are naturally superimposed. The system reconstruction quality decreases smoothly as the channel noise increases. This fundamentally overcomes the "cliff effect" of traditional modular encryption schemes that trigger catastrophic decoding or decryption failures due to a slight increase in channel error under limited code length or time-varying channel conditions. It achieves the unity of cryptographic security and the graceful degradation characteristics of semantic communication.

[0018] (2) This invention constructs a gradient bypass channel that bypasses non-differentiable modules such as encryption, modulation, channel transmission, demodulation and decryption by setting a first pass-through estimator node at the output of the quantizer at the transmitting end and a second pass-through estimator node at the input of the dequantizer at the receiving end. This enables the error signal at the receiving end to propagate back to the semantic encoder at the transmitting end. For the first time, end-to-end gradient backpropagation and joint optimization are realized in the encrypted semantic communication system. This solves the technical problem that the encoder cannot learn a robust semantic representation of encryption-channel composite noise because the homomorphic encryption pipeline is completely non-differentiable. (3) The present invention utilizes the gradient bypass channel established by the encrypted gradient tunneling mechanism, takes the error between the reconstructed image and the input image as the loss function, performs end-to-end joint fine-tuning of the semantic encoder and semantic decoder, and adopts step linear heating scheduling to control the soft allocation temperature parameter of the quantizer during the fine-tuning process, so that the soft allocation provides a stable gradient in the early stage of training and gradually approaches hard quantization in the later stage of training to eliminate the inconsistency between training and inference, so that the encoder can actively learn to adapt to the combined disturbance of encrypted noise and channel noise, and obtain significant performance gains under low signal-to-noise ratio. (4) After hard decision demodulation at the receiving end, the present invention measures the bit error rate of the current transmission block in real time and compares it with the pre-calibrated threshold value. Based on the comparison result, it adaptively selects the robust decoding model after joint encryption fine-tuning or the pre-trained high-fidelity decoding model for image reconstruction. This mechanism effectively eliminates the influence of the small precision loss introduced by encryption under high signal-to-noise ratio on the reconstruction fidelity, so that the encryption system can achieve reconstruction quality that is almost indistinguishable from the unencrypted baseline under high signal-to-noise ratio, thus taking into account both security and high fidelity. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the semantic image transmission method based on homomorphic encryption and deep joint source channel coding of the present invention.

[0021] Figure 2 This is a comparison chart of the reconstructed image quality of the method of the present invention and existing solutions under different SNR conditions.

[0022] Figure 3 This is a graph comparing the peak signal-to-noise ratio performance of the method of the present invention with that of existing solutions.

[0023] Figure 4This diagram illustrates the reconstruction effect and security gap of the method of the present invention under eavesdropping attacks. Detailed Implementation

[0024] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] This embodiment takes a wireless image transmission system as an example. Both the transmitting and receiving ends are equipped with a deep joint source-channel coding model based on SwingTransformer. The system bandwidth ratio is... ρ =0.3333.

[0026] Perform the following initialization configurations before system deployment: The receiving end (Bob) generates a BFV homomorphic encryption private key s and a public key P(s). The parameters are configured as follows: plaintext modulus p = 256, ciphertext modulus q = 65536, polynomial degree N = 4096, and the error distribution is a discrete Gaussian distribution with a standard deviation of [missing information]. This corresponds to a 128-bit security level. The public key is publicly disclosed to all senders, while the private key is secretly kept by the receiver.

[0027] The semantic encoder and decoder are pre-trained on the DIV2K dataset. During pre-training, the channel signal-to-noise ratio is uniformly and randomly sampled within the range of 1dB to 30dB, using an AWGN channel model and a mean squared error loss function. After pre-training, the minimum value of the global semantic features is saved. and maximum value .

[0028] Reference Figure 1 This invention provides a semantic image transmission method based on homomorphic encryption and deep joint source-channel coding, comprising the following steps: Step S1, Semantic Encoding and Quantization The sending end extracts the semantic latent representation from the input image using a semantic encoder, and then linearly scales and quantizes the semantic latent representation to obtain integer plaintext; specifically, it includes the following steps: Sending end obtains input image Semantic encoder A hierarchical Swing Transformer structure is adopted, comprising four stages. The first stage divides the image into 2×2 non-overlapping blocks and linearly projects them into C_1 dimensional semantic tokens, which are then processed through N_1 Swing Transformer blocks (alternating between window multi-head self-attention and shift window multi-head self-attention). The second to fourth stages perform block merging (concatenating adjacent 2×2 tokens along the channels and projecting them to reduce dimensionality, doubling the number of channels and halving the spatial resolution), followed by processing through several Swing Transformer blocks. The fourth stage outputs a semantic feature map. .

[0029] After the fourth stage, a learnable 1×1 convolutional layer is connected to compress the number of channels from 320 to C, resulting in the final semantic latent representation. In this embodiment, C is based on the bandwidth ratio ρ =0.3333 is confirmed.

[0030] Using the global minimum determined during the pre-training phase and maximum value semantic latent representation Linear scaling to the [0, 255] interval yields continuously scaled values. The scaling formula is: ; in, and These are the semantic latent representations before and after scaling, respectively; At the first pass-through estimation node (STE-1) of the Encrypted Gradient Tunneling (EGT) mechanism, hard rounding is performed during forward propagation. , This represents the semantic latent representation after hard rounding, i.e., the plaintext integer; Backpropagation employs distance-based soft-quantized surrogate gradients for scaling values. and centroid set The formula for calculating soft quantization weights is: , ; in, For temperature parameters, Scaling value With the center of mass The square of the Euclidean distance; Softening value The calculation formula is: ; The gradient of the soft path is backpropagated to the encoder as a gradient estimate with hard rounding.

[0031] Step S2, BFV homomorphic encryption

[0032] The sending end uses the BFV homomorphic encryption scheme to encrypt the integer plaintext, generating semantic ciphertext; specifically, it includes the following steps: The quantized integer plaintext sequence Input the BFV encryption module, for each plaintext integer Encode it into a polynomial using the encryption function E. The constant term is The remaining coefficients are zero; the sampling noise vector is , , Calculate the ciphertext pair: ; ; in, and For encrypted pairs, For the common matrix, For private key, To and The corresponding public key component, The error distribution is discrete Gaussian. For the ciphertext modulus, and Let be the dimension of the noise vector. is the length of the plaintext vector.

[0033] Step S3, Channel Modulation and Transmission

[0034] The transmitting end converts the semantic ciphertext into a binary sequence, performs Gray-coded QAM modulation, generates baseband symbols, and transmits them through the wireless channel; specifically, it includes the following steps: Each ciphertext integer Encoded as a 16-bit binary sequence (padded with zeros at the high bits): ; The binary stream is grouped into 6-bit blocks, with zeros padded if the last block is less than 6 bits. Each block is mapped to 64-QAM constellation points using Gray coding, and the modulated symbol sequence y is transmitted via a wireless channel.

[0035] In an AWGN (Additive White Gaussian Noise) channel, the received signal is: ; In a flat Rayleigh fading channel, the received signal is: .

[0036] in, This is the signal vector received by the receiver. The baseband symbol vector transmitted by the transmitting end; It is an additive complex Gaussian white noise vector. It follows a complex Gaussian distribution with a mean of 0 and a covariance matrix of . Multiply by the identity matrix ,in For noise variance; Let be the Rayleigh fading channel coefficient vector, where each element is independent and identically distributed, following a complex Gaussian distribution with mean 0 and variance 1. .

[0037] Step S4: Receive, demodulate, and recover semantic ciphertext

[0038] The receiving end performs hard-decision demodulation on the received signal to recover the noisy semantic ciphertext. The specific method is as follows: The receiving end performs hard-decision demodulation on the received signal: ; in, The QAM constellation points are estimated after hard decision demodulation. For the received first A symbol, For the baseband symbol to be transmitted, For the 64-QAM constellation point set; The receiving end will Mapping back to Gray code tags, removing padding bits, and recombining them into 16-bit words yields the semantically ciphertext containing noise. .

[0039] Step S5, BFV decryption and inverse quantization output

[0040] The receiving end uses the BFV private key to decrypt the noisy semantic ciphertext, obtaining integer plaintext that is jointly perturbed by channel error and BFV decryption noise. This plaintext is then dequantized to recover the continuous semantic representation. Specifically, this includes the following steps: The receiving end uses the private key s pair BFV decryption yields integer plaintext affected by a combination of channel error and decryption noise. ; At the second pass-through estimator node (STE-2) of the EGT (Encrypted Gradient Tunneling) mechanism, for the decrypted integer plaintext... Calculate its relationship with each centroid. square Euclidean distance : ; The soft-assignment weights are obtained by temperature-weighted softmax. : , ; in, For temperature parameters; Soft dequantization outputs continuous values for: ; During backpropagation, the gradient is propagated back using the following formula: ; gradient The system is routed directly to the first straight-through estimator node (STE-1) via the identity mapping link, bypassing the non-differentiable module.

[0041] Step S6, Semantic Decoding and Image Reconstruction

[0042] The receiving end reconstructs the continuous semantic representation into an output image using a semantic decoder, specifically including the following steps: Will After inverse scaling (restoring to the original semantic representation range), it is input into the semantic decoder. The decoder employs a Swing Transformer structure symmetrical to the encoder, progressively restoring the feature maps to their original resolution through four upsampling stages, ultimately outputting the reconstructed image. The loss function uses mean squared error: .

[0043] Specifically, the method also includes a joint encryption fine-tuning strategy: A first pass-through estimator node is set at the output of the quantizer at the transmitting end, and a second pass-through estimator node is set at the input of the dequantizer at the receiving end. This constructs a gradient bypass channel that bypasses encryption, modulation, channel transmission, demodulation, and decryption steps. Using this gradient bypass channel, the error between the reconstructed image and the input image is used as the loss function, and the parameters of the semantic encoder and decoder are used as optimization variables. End-to-end joint fine-tuning of the encoder and decoder is performed by minimizing this loss function. The fine-tuning objective function is expressed as: ; in, and These are the parameters for the encoder and decoder, respectively. For the input image, The channel signal-to-noise ratio (SNR) For loss function, For decoder, For semantic representations affected by compound noise; uniform sampling within the range of 1dB to 30dB.

[0044] During the joint encryption fine-tuning process, the soft-assignment temperature parameter of the quantizer increases from its initial value to its maximum value according to a step-linear temperature ramping schedule, which is expressed as follows: ; in, This represents the current number of training steps. This represents the total number of steps. The starting temperature, This is the scaling factor. The maximum temperature, This indicates rounding down to the nearest integer. In this embodiment, , , , In the early stages of training, a wide soft gradient allocation provides a stable gradient, which is then gradually approximated to hard quantization to eliminate training-inference inconsistencies. Fine-tuning is performed for 200 epochs with a batch size of 16, using the Adam optimizer and a learning rate of... .

[0045] Preferably, the method further includes a bit error rate gating inference mechanism: The receiver measures the bit error rate of the current transport block in real time after the hard decision demodulation in step S4. Bit error rate With pre-calibrated threshold value Compare; If the bit error rate is higher than the threshold, then the robust decoding model finely tuned by joint encryption is selected for reconstruction; If the bit error rate is not higher than the threshold, then a pre-trained high-fidelity decoding model is selected for reconstruction.

[0046] Furthermore, the threshold value is calibrated in the following manner: At a set of signal-to-noise ratio sampling points (e.g., 1, 7, 13, 19, 25, 30 dB), the average signal-to-noise ratio of the pre-trained model was recorded. and the average of the fine-tuning model and the average experience of pre-trained models Computational performance advantages ,in To provide a conservative margin, the default value is 0; find the crossover interval where the performance advantage of the fine-tuned model relative to the pre-trained model changes from positive to negative. and The optimal bit error rate threshold is calculated using linear interpolation within the aforementioned cross-interval: ; If all ,but (Always use pre-trained models); if all ,but (Always use the fine-tuned model).

[0047] In this embodiment, the cryptographic algorithm used is the SM2 / SM3 / SM4 algorithm issued by the State Cryptography Administration or a homomorphic encryption algorithm that conforms to the RLWE problem; when the national cryptographic algorithm is used, the polynomial ring operation in the BFV homomorphic encryption is replaced with a variant based on SM2 elliptic curve cryptography.

[0048] This embodiment also evaluates the security against the eavesdropper Eve. It is assumed that Eve possesses a perfect channel, a public key, a deep learning model of the same capacity as Bob's, and fully trained model parameters. Under a ciphertext-only attack strategy, Eve attempts to directly decrypt the ciphertext, fine-tune the neural network to approximate the decryption function, or make random guesses. Experimental results show that Eve's reconstructed PSNR does not exceed 11.73 dB under all attacks, and the SSIM is between -0.0609 and 0.1117, failing to recover any valid image structure. Meanwhile, the legitimate receiver Bob's PSNR increases from 16.70 dB at low SNR to 33.11 dB at high SNR, and the security gap (the difference in PSNR between Bob and Eve) expands from 5.28 dB to 21.95 dB, confirming the anti-eavesdropping capability of this invention.

[0049] Figure 2 To compare the visual effects of different schemes under different SNRs, this figure uses a starfish image as an example to show the visual reconstruction effect of the SecureWITT invention and all comparison schemes under six typical signal-to-noise ratios of 1dB, 7dB, 13dB, 19dB, 25dB, and 30dB. The PSNR value and semantic recognizability rate are marked below each sub-figure.

[0050] Core overall effect conclusion: Smooth and graceful degradation across the entire SNR range: SecureWITT exhibits continuous performance degradation characteristics across all SNR levels: at 1dB, the basic outline of a starfish is recognizable (PSNR 18.1dB, recognizable rate 3%); at 7dB, the outline is clearly discernible (20.4dB, recognizable rate 10%); at 13dB, the details are basically intact (22.1dB, recognizable rate 36%); at 19dB, it approaches the quality of an unencrypted baseline (27.9dB, recognizable rate 83%); and at 25~30dB, it is visually almost indistinguishable from an unencrypted WITT baseline (32.8dB, recognizable rate over 95%).

[0051] The performance of the comparison encryption schemes is completely ineffective: all comparison schemes with direct concatenated homomorphic encryption are completely unable to reconstruct effective images in the 1~19dB range (PSNR<12dB, recognition rate 0%); only above 25dB do blurred outlines begin to appear, but the reconstruction quality is far lower than that of the scheme of this invention.

[0052] The only solution to achieve a balance between security and performance: This invention is currently the only secure semantic image transmission solution that simultaneously meets the three core requirements of "architectural-level 128-bit cryptographic security + smooth and graceful degradation across the entire SNR range + performance comparable to the unencrypted baseline under high signal-to-noise ratio". It is suitable for scenarios with strict requirements for both security and robustness, such as UAV backhaul, telemedicine, and military reconnaissance.

[0053] Figure 3 The graph shows a PSNR vs SNR performance comparison curve. The horizontal axis represents the channel signal-to-noise ratio (SNR, 1~30dB), and the vertical axis represents the peak signal-to-noise ratio of the reconstructed image (PSNR, dB, the higher the value, the better the reconstruction quality). It compares the end-to-end transmission performance of the SecureWITT scheme proposed in this invention with unencrypted baselines (WITT, DeepJSCC) and other direct concatenated homomorphic encryption schemes (WITT-RLWE, WITT-MLWE, DeepJSCC-BFV, etc.).

[0054] Key experimental conclusions: The performance of SecureWITT across the entire SNR range is far superior to other encryption schemes: all comparison schemes that directly cascade homomorphic encryption with JSCC have PSNR below 12dB in the 1~25dB range, making it completely impossible to reconstruct a valid image; while the present invention SecureWITT maintains usable reconstruction quality across the entire SNR range.

[0055] Significantly improved robustness at low signal-to-noise ratios: In the 1-13dB low signal-to-noise ratio range, SecureWITT's PSNR linearly increases from 17.3dB to 24.2dB, exhibiting perfectly smooth and elegant degradation characteristics, completely overcoming the "cliff effect" of traditional modular encryption; compared with the scheme that only fine-tunes the decoder, it achieves a PSNR gain of 3-5dB, verifying the effectiveness of the Encryption Gradient Tunneling (EGT) and Joint Encryption Fine-tuning (JEF) mechanisms in enabling the encoder to learn encryption-aware robust representations.

[0056] Negligible fidelity loss at high signal-to-noise ratios: In the 25-30dB high signal-to-noise ratio range, SecureWITT's PSNR quickly converges to 35.8dB, narrowing the gap with the unencrypted WITT baseline to within 0.013dB. This demonstrates that the bit error rate gated inference (BGI) mechanism successfully eliminates the slight precision loss introduced by encryption, achieving a perfect balance between security and high fidelity.

[0057] Figure 4This is a visual comparison of a legitimate receiver and an eavesdropper at a 13dB SNR. Under the condition that the channel signal-to-noise ratio is fixed at 13dB, this image compares the reconstruction results of the original image, the eavesdropper Eve without encryption, the reconstruction results of Eve under two attacks (direct decryption and fine-tuning model optimal attack) using the scheme of this invention, and the reconstruction result of the legitimate receiver Bob. The corresponding PSNR and structural similarity (SSIM) index are labeled below each sub-image.

[0058] Core security and performance conclusions: The unencrypted scheme poses a serious risk of eavesdropping: without encryption, the reconstructed PSNR of the eavesdropper Eve is as high as 32~41dB, and the SSIM is close to 1, which is almost indistinguishable from the original image, proving that the traditional Deep JSCC scheme has no cryptographic security protection at all.

[0059] This invention possesses extremely strong anti-eavesdropping capabilities: even if an eavesdropper has a perfect channel, public key, and fully trained model parameters, the PSNR of the direct decryption reconstruction is only 10.5~12.7dB, and the SSIM is close to 0, which is completely random noise texture; even if Eve performs targeted fine-tuning and optimization for the ciphertext, the optimal reconstruction PSNR is still only 10.3~12.4dB, and the highest SSIM is only 0.19, which still cannot recover any effective semantic content.

[0060] The reconstruction quality of the legitimate receiver is excellent: the reconstructed PSNR of the legitimate receiver Bob reaches 22.9~25.1dB, the SSIM is 0.93~0.98, the image is clear and semantically complete, and the visual difference from the original image is minimal; the security gap (PSNR difference) between the legitimate receiver and the eavesdropper reaches 11~15dB, and further increases with the increase of the channel signal-to-noise ratio.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A semantic image transmission method based on homomorphic encryption and deep joint source-channel coding, characterized in that, Includes the following steps: S1, the sending end extracts the input image into a semantic latent representation through a semantic encoder, and performs linear scaling and quantization on the semantic latent representation to obtain integer plaintext; S2, the sending end uses the BFV homomorphic encryption scheme to encrypt the integer plaintext and generate semantic ciphertext; S3, the transmitting end converts the semantic ciphertext into a binary sequence, performs Gray-coded QAM modulation, generates baseband symbols, and transmits them through the wireless channel; S4, the receiving end performs hard decision demodulation on the received signal to recover the noisy semantic ciphertext; S5, the receiver uses the BFV private key to decrypt the noisy semantic ciphertext, obtaining the integer plaintext that is jointly disturbed by channel error and BFV decryption noise, and then recovers the continuous semantic representation through inverse quantization; S6, the receiving end reconstructs the continuous semantic representation into an output image through a semantic decoder.

2. The semantic image transmission method based on homomorphic encryption and deep joint source-channel coding as described in claim 1, characterized in that, The method also includes an encrypted gradient tunneling mechanism: A first pass-through estimator node is set at the output of the quantizer at the transmitting end, and a second pass-through estimator node is set at the input of the dequantizer at the receiving end. A gradient bypass channel that bypasses the encryption, modulation, channel transmission, demodulation and decryption steps is constructed to backpropagate the error signal at the receiving end to the semantic encoder at the transmitting end, so as to achieve end-to-end joint optimization.

3. The semantic image transmission method based on homomorphic encryption and deep joint source-channel coding as described in claim 2, characterized in that, The method also includes a joint encryption fine-tuning strategy: Using the gradient bypass channel, the error between the reconstructed image and the input image is used as the loss function, and the parameters of the semantic encoder and semantic decoder are used as optimization variables. The encoder and decoder are jointly fine-tuned end-to-end by minimizing the loss function. The fine-tuning objective function is expressed as: ； in, and These are the parameters for the encoder and decoder, respectively. For the input image, The channel signal-to-noise ratio (SNR) For loss function, For decoder, This is a semantic representation affected by complex noise. During the joint encryption fine-tuning process, the soft-assignment temperature parameter of the quantizer increases from its initial value to its maximum value according to a step-linear temperature ramping schedule, which is expressed as follows: ； in, This represents the current number of training steps. This represents the total number of steps. The starting temperature, This is the scaling factor. The maximum temperature, This indicates rounding down to the nearest integer.

4. The semantic image transmission method based on homomorphic encryption and deep joint source-channel coding as described in claim 3, characterized in that, The method also includes a bit error rate gated inference mechanism: After hard decision demodulation, the receiver measures the bit error rate of the current transmission block in real time and compares the bit error rate with a pre-calibrated threshold value. If the bit error rate is higher than the threshold, the model fine-tuned by joint encryption is selected for reconstruction; If the bit error rate is not higher than the threshold, then the pre-trained model is selected for reconstruction.

5. The semantic image transmission method based on homomorphic encryption and deep joint source-channel coding as described in claim 4, characterized in that, The threshold value is calibrated in the following manner: The average peak signal-to-noise ratio of the pre-trained model and the fine-tuned model, as well as the average bit error rate experienced by the pre-trained model, are recorded at a set of signal-to-noise ratio sampling points to calculate the performance advantage. Find the crossover interval where the performance advantage of the fine-tuned model relative to the pre-trained model turns from positive to negative, and calculate the optimal bit error rate threshold within the crossover interval using linear interpolation.

6. The semantic image transmission method based on homomorphic encryption and deep joint source-channel coding as described in claim 2, characterized in that, In step S1, the quantization operation specifically involves: Using the global minimum determined during the pre-training phase and maximum value The semantic latent representation is linearly scaled to the [0, 255] interval, and the scaling formula is: ； in, and These are the semantic latent representations after scaling and before scaling, respectively; Then perform the rounding operation: At the first straight-through estimation node of the encrypted gradient tunneling mechanism, hard rounding is performed during forward propagation. , This represents the semantic latent representation after hard rounding, i.e., the plaintext integer; Backpropagation employs distance-based soft-quantization surrogate gradients, where the soft-quantization weights are calculated using the following formula: ， ； in, For temperature parameters, Scaling value With center of mass The square of the Euclidean distance; Softening value The calculation formula is: ； The gradient of the soft path is backpropagated to the encoder as a gradient estimate with hard rounding.

7. The semantic image transmission method based on homomorphic encryption and deep joint source-channel coding as described in claim 1, characterized in that, In step S2, the ciphertext generation method for BFV homomorphic encryption is as follows: For each plaintext integer Encoded as a polynomial The constant term is The remaining coefficients are zero; the sampling noise vector is , , Calculate the ciphertext pair: ； ； in, and For encrypted pairs, For the common matrix, For private key, To and The corresponding public key component, The error distribution is discrete Gaussian. For the ciphertext modulus, and Let be the dimension of the noise vector. The length of the plaintext vector.

8. The semantic image transmission method based on homomorphic encryption and deep joint source-channel coding as described in claim 1, characterized in that, In step S4, the hard decision demodulation rule is: ； in, The QAM constellation points are estimated after hard decision demodulation. For the received first A symbol, For the baseband symbol to be transmitted, For the 64-QAM constellation point set; The receiving end will Mapping back to Gray code labels, removing padding bits, and then reassembling yields semantically ciphertext containing noise.

9. A semantic image transmission method based on homomorphic encryption and deep joint source-channel coding as described in claim 2, characterized in that, In step S5, the inverse quantization employs a software dequantization operation: For the decrypted plaintext integer Calculate its relationship with each centroid. square Euclidean distance : ； The soft-assigned weights are obtained by temperature-weighted softmax. : ， ； in, For temperature parameters; Soft dequantization outputs continuous values for: ； During backpropagation, the gradient is propagated back using the following formula: ； gradient It is directly routed to the first direct-access estimator node via the identity mapping link.

10. A semantic image transmission method based on homomorphic encryption and deep joint source-channel coding as described in claim 1, characterized in that, The semantic encoder adopts a hierarchical Swing Transformer structure, which includes four stages. Each stage performs spatial downsampling and semantic feature transformation in sequence. After the last stage, a learnable 1×1 convolutional layer is connected to compress the number of channels of the feature map to the dimension corresponding to the target semantic channel bandwidth ratio. The semantic decoder adopts a Swing Transformer structure symmetrical to the semantic encoder, which gradually restores the feature map to the original resolution through four upsampling stages.

Images