From Zero to Hero: Cross-modal-enhanced Adversarial Item Promotion Attack against Multimodal Recommender Systems

Mengyu Yao, MOEKey Lab of HCST (PKU), School of Computer Science, Peking University; Ant Group; Ziqi Zhang, University of Illinois Urbana-Champaign; Yifeng Cai and Junlin Liu, MOEKey Lab of HCST (PKU), School of Computer Science, Peking University; Xinyi Fu and Weiqiang Wang, Ant Group; Xiangqun Chen, Yao Guo, and Ding Li, MOEKey Lab of HCST (PKU), School of Computer Science, Peking University

Multimodal recommender systems (MRSs) jointly leverage visual and textual item representations to determine item ranking and exposure in modern online platforms. Their strong dependence on item content, however, introduces new security risks. In particular, malicious sellers can conduct the adversarial item promotion (AIP) attack to manipulate item content to deceive the recommender into overranking specific items. When it succeeds, the promoted items gain disproportionately higher exposure, leading to significant market visibility and direct economic gain. However, existing AIP research primarily targets unimodal recommenders, leaving the unique vulnerabilities of MRSs largely unexplored. Meanwhile, multimodal adversarial attacks for vision–language models (VLMs) optimize objectives unrelated to ranking and lack cross-modal coordination unique to MRSs. To bridge this gap, we propose CREAM (CRoss-modal-Enhanced AIP attack against MRSs). Our key insight is to jointly perturb multiple modalities in a semantically consistent manner}. We integrate a tailored visual perturbator, a text generator, and a joint optimization controller to fully exploit cross-modal correlations in a black-box setting. Our comprehensive evaluation shows that CREAM significantly outperforms existing methods, achieving on average 5.75x and 2.89x higher exposure at top-10 and top-50 metrics, and demonstrates robustness under evolving real-world conditions. This exposes a tangible economic risk to recommender platforms. Meanwhile, CREAM maintains high imperceptibility across visual, textual, and cross-modal dimensions. We further investigate several potential defense strategies and demonstrate their limitations, highlighting the urgent need for stronger protections against adversarial threats in MRSs.

USENIX Security '26 Open Access Sponsored by
King Abdullah University of Science and Technology (KAUST)

Open Access Media

USENIX is committed to Open Access to the research presented at our events. Papers and proceedings are freely available to everyone once the event begins. Any video, audio, and/or slides that are posted after the event are also free and open to everyone. Support USENIX and our commitment to Open Access.

BibTeX
@inproceedings {320685,
author = {Mengyu Yao and Ziqi Zhang and Yifeng Cai and Junlin Liu and Xinyi Fu and Weiqiang Wang and Xiangqun Chen and Yao Guo and Ding Li},
title = {From Zero to Hero: Cross-modal-enhanced Adversarial Item Promotion Attack against Multimodal Recommender Systems},
booktitle = {35th USENIX Security Symposium (USENIX Security 26)},
year = {2026},
isbn = {978-1-939133-58-8},
address = {Baltimore, MD},
pages = {4781--4800},
url = {https://www.usenix.org/conference/usenixsecurity26/presentation/yao-zero},
publisher = {USENIX Association},
month = aug
}