From Internal Diagnosis to External Auditing: A VLM-Driven Paradigm for Data-Free Online Backdoor Defense
Researchers have introduced PRISM, a novel framework for defending against backdoor attacks on deep neural networks. This approach shifts from internal model diagnosis to external semantic auditing, utilizing Universal Vision-Language Models (VLMs) as independent security auditors. PRISM refines visual prototypes online and uses an adaptive router for real-time threshold calibration, demonstrating state-of-the-art performance in suppressing attack success rates while maintaining clean accuracy. AI
IMPACT Introduces a novel, externalized defense mechanism against backdoor attacks, potentially enhancing the security of deployed AI models.