ML Security PapersFrom Internal Diagnosis to External Auditing: A VLM-Driven Paradigm for Online Test-Time Backdoor Defense
Binyan Xu 1, Fan Yang 1, Xilin Dai 2, Di Tang 3, Kehuan Zhang 1
Published on arXiv
2601.19448
Model Poisoning
OWASP ML Top 10 — ML10
Key Finding
Achieves Attack Success Rate below 1% on CIFAR-10 across 11 backdoor attack types while improving clean accuracy, outperforming existing test-time defenses.
PRISM (Prototype Refinement & Inspection via Statistical Monitoring)
Novel technique introduced
Deep Neural Networks remain inherently vulnerable to backdoor attacks. Traditional test-time defenses largely operate under the paradigm of internal diagnosis methods like model repairing or input robustness, yet these approaches are often fragile under advanced attacks as they remain entangled with the victim model's corrupted parameters. We propose a paradigm shift from Internal Diagnosis to External Semantic Auditing, arguing that effective defense requires decoupling safety from the victim model via an independent, semantically grounded auditor. To this end, we present a framework harnessing Universal Vision-Language Models (VLMs) as evolving semantic gatekeepers. We introduce PRISM (Prototype Refinement & Inspection via Statistical Monitoring), which overcomes the domain gap of general VLMs through two key mechanisms: a Hybrid VLM Teacher that dynamically refines visual prototypes online, and an Adaptive Router powered by statistical margin monitoring to calibrate gating thresholds in real-time. Extensive evaluation across 17 datasets and 11 attack types demonstrates that PRISM achieves state-of-the-art performance, suppressing Attack Success Rate to <1% on CIFAR-10 while improving clean accuracy, establishing a new standard for model-agnostic, externalized security.
Key Contributions
- Paradigm shift from internal model diagnosis to external semantic auditing, decoupling backdoor defense from the victim model's corrupted parameters
- PRISM framework featuring a Hybrid VLM Teacher for online visual prototype refinement to bridge the domain gap of general-purpose VLMs
- Adaptive Router with statistical margin monitoring to calibrate gating thresholds in real-time across 17 datasets and 11 attack types
🛡️ Threat Analysis
PRISM is explicitly designed as a test-time defense against backdoor/trojan attacks — it detects trigger-activated inputs during inference using an external VLM auditor, directly addressing the backdoor threat model across 11 attack types.