Enhancing Image Privacy in Semantic Communication over Wiretap Channels leveraging Differential Privacy

TL;DR

This paper proposes a novel semantic communication framework that employs differential privacy and GAN techniques to protect image privacy over wiretap channels, ensuring secure and high-fidelity transmission.

Contribution

It introduces a differential privacy-based image protection mechanism integrated with GAN inversion and neural networks to enhance privacy in semantic communication.

Findings

Effective privacy protection against eavesdroppers.

High-fidelity image reconstruction at legitimate receivers.

Near-original differential privacy levels achieved.

Abstract

Semantic communication (SemCom) enhances transmission efficiency by sending only task-relevant information compared to traditional methods. However, transmitting semantic-rich data over insecure or public channels poses security and privacy risks. This paper addresses the privacy problem of transmitting images over wiretap channels and proposes a novel SemCom approach ensuring privacy through a differential privacy (DP)-based image protection and deprotection mechanism. The method utilizes the GAN inversion technique to extract disentangled semantic features and applies a DP mechanism to protect sensitive features within the extracted semantic information. To address the non-invertibility of DP, we introduce two neural networks to approximate the DP application and removal processes, offering a privacy protection level close to that by the original DP process. Simulation results validate the effectiveness of our method in preventing eavesdroppers from obtaining sensitive information while maintaining high-fidelity image reconstruction at the legitimate receiver.