FlowGuard Detects Multimodal Attacks via Internal Information Decomposition

FlowGuard is a lightweight inference-time framework that monitors cross-modal consistency to detect adversarial attacks. It uses FlowVectors to identify when malicious intent is distributed across text and vision modalities to evade unimodal safeguards.