UAT-MC: New Defense Against Evasion-Based Promotion Attacks in Multimodal Recommender Systems

In a recent study, researchers discovered a cross-modal gradient mismatch in multimodal recommender systems during multi-user promotion scenarios. They found that visual and textual perturbations are optimized in conflicting directions because of prevailing user groups, which weakens the effectiveness of attacks and leads to an underestimation of worst-case risks during robust training. To tackle this issue, they introduce Untargeted Adversarial Training with Multimodal Coordination (UAT-MC), which considers every item as a possible target to defend against evasion-based attacks. The findings are documented in arXiv:2605.06238.