Multimodal large language models (MLLMs) are vulnerable to backdoor attacks, but most defenses inspect static snapshots of model internals or look for specific trigger patterns. The authors of a new paper argue that this approach is too rigid: textual injections, sub-perceptual noise, and coordinated cross-modal triggers can easily evade such priors, especially when the attack type is unknown.
Their proposed framework, EVO-Detect, shifts the focus to dynamic output evolution. The key insight is that a successful backdoor hijacks the model's decoding trajectory, skewing how image and text modalities contribute during generation while keeping outputs abnormally invariant to input perturbations. EVO-Detect captures this through two complementary channels: an attention flow branch that tracks how image-text attention allocation changes during early generation, and a logits flow branch that measures output distribution sensitivity to modal perturbations via interventional ablations.
The trajectories are converted into compact sample-level metrics and normalized against negative-control signatures, reducing false alarms from natural perturbations. In evaluations, EVO-Detect substantially outperformed existing baselines BYE and TCAP across multiple attack families. The paper was accepted as an oral presentation at ACM MM 2026.