Privacy-Preserving Semantic Communication over Wiretap Channels with Learnable Differential Privacy

A team from Zhejiang University and the University of Technology Sydney has published a paper on arXiv proposing a novel secure semantic communication framework for image transmission over wiretap channels. The work, titled 'Privacy-Preserving Semantic Communication over Wiretap Channels with Learnable Differential Privacy,' addresses critical privacy vulnerabilities in semantic communication (SemCom) systems. Traditional secure SemCom often relies on impractical assumptions like favorable channel conditions or prior knowledge of eavesdropper models. The new framework uses differential privacy (DP) with learnable noise patterns, generated through adversarial training of neural networks, instead of conventional white Gaussian or Laplace noise. This design mitigates the non-invertibility problem of DP while providing approximate privacy guarantees and controllable security levels via adjustable privacy budgets.

Experimental results show the method significantly degrades reconstruction quality for eavesdroppers while introducing only slight degradation in task performance for legitimate users. Compared to prior DP-based methods, the approach achieves a Learned Perceptual Image Patch Similarity (LPIPS) advantage of 0.06-0.29 and a Fréchet Perceptual Patch Similarity Ratio (FPPSR) advantage of 0.10-0.86 under comparable security levels. The framework first extracts disentangled semantic representations from source images using a generative adversarial network (GAN) inversion method, then selectively perturbs private semantic representations with approximate DP noise. This work represents a significant step toward practical, privacy-preserving semantic communication systems that can operate without restrictive assumptions about channel conditions or eavesdropper capabilities.