Secure Mobile Edge Generation (MEG): Privacy Preservation in the Presence of Eavesdroppers

Abstract

Mobile edge generation (MEG) enables low-latency AI-generated content (AIGC) by decentralizing large generative AI (GAI) models across the edge server (ES) and user equipment (UE). However, the distributed collaborative generation raises significant privacy concerns under potential communication eavesdropping. To address this issue, a novel secure MEG framework is proposed for image generation, which integrates a customized privacy-aware encoder-decoder module into the decentralized GAI architecture. Then, latent diffusion features can be securely transmitted from the ES to the UE via joint source-channel coding (JSCC), which supports cost-efficient and privacy-preserving ES-UE collaboration. To assess privacy leakage, a new metric termed privacy-preserving structural similarity (PP-SSIM) is introduced. The proposed modules are jointly trained in an end-to-end manner to optimize both reconstruction fidelity and privacy preservation. Experimental results show that the proposed method achieves low-latency and secure image generation for legitimate users, while effectively mitigating information leakage to eavesdroppers. Despite a slight degradation in visual quality, the perceptual fidelity is well preserved.