SemCom-UAP: Universal Adversarial Perturbations Against Cross-Modal Retrieval in Semantic Communications

Abstract

Semantic communication (SemCom) systems based on vision-language pre-training (VLP) models transmit compact embeddings for downstream tasks such as cross-modal retrieval, but also introduce new vulnerabilities to universal adversarial perturbations (UAPs). Existing VLP-targeted UAP methods typically assume noise-free transmission and therefore fail to account for channel corruption encountered in realistic SemCom deployments. In this paper, we propose SemCom-UAP, a channel-aware UAP generation framework for cross-modal retrieval in SemCom systems that jointly incorporates cross-modal semantics and wireless channel state into signal-to-noise ratio (SNR)-aware perturbation generators. To improve attack robustness under varying wireless conditions, we further develop a channel-robust adversarial optimization strategy that integrates adversarial cross-modal contrastive learning on additive white Gaussian noise (AWGN)-corrupted embeddings, adaptive multi-surrogate training, and stratified SNR sampling. Experiments on Flickr30K with six VLP backbones show that SemCom-UAP consistently outperforms channel-blind baselines across diverse SNR conditions, highlighting the practical adversarial threat to SemCom retrieval systems.