Research map / Adversarial robustness and security / Backdoor and poisoning
From Internal Diagnosis to External Auditing: A VLM-Driven Paradigm for Data-Free Online Backdoor Defense
Binyan Xu, Fan Yang, Xilin Dai, Di Tang, Kehuan Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PRISM external VLM auditing instead of internal-diagnosis defenses | data-free backdoor defenses · CIFAR-10 (average over 11 attacks) | ASR: Refine 14.3 -> PRISM 0.8 | Improved |
| use PRISM instead of zero-shot VLM classification | VLM auditor · CIFAR-10 (average over 11 attacks) | clean accuracy: 86.7 -> 93.2 | Improved |
| average victim and VLM logits (Ensemble) instead of PRISM routing | victim + VLM · CIFAR-10 (average over 11 attacks) | ASR: 0.8 -> 75.7 | Hurt |
| add online update to PRISM (ablation) | PRISM · GTSRB, classic attacks | ASR: 32.9 -> 1.3 | Improved |
| add prototype refinement to PRISM (ablation) | PRISM · GTSRB, classic attacks | clean accuracy: 86.1 -> 94.4 | Improved |
| add Cornish-Fisher skewness correction to the adaptive threshold (ablation) | PRISM · GTSRB, classic attacks | ASR and CA: ASR 0.8 (w/o) vs 1.3 (full) | No clear effect |
| add skewness correction so a single threshold zeta=-2 works across datasets (ablation) | PRISM · CIFAR and GTSRB threshold sensitivity | optimal threshold consistency | Improved |
| typographic adaptive prompt attack superimposing target class name (ablation) | PRISM · CIFAR-10 | ASR: 1.3 -> 15.6 | Hurt |
| backdoor the auditing VLM with the same trigger and target as the victim (ablation) | PRISM · CIFAR-10, BadNets | ASR: 0.2 -> 100 | Hurt |
| backdoor the auditing VLM with a different trigger or different target (ablation) | PRISM · CIFAR-10, BadNets | ASR: 0.2 -> 0.8 (different trigger), 0.0 (different target) | No clear effect |
| flooding, periodic and mixed online adaptive attacks (ablation) | PRISM · GTSRB online stream | ASR | No clear effect |
| use poison set 50x larger than clean set (ablation) | PRISM · dataset size scalability | ASR: ~12% | Hurt |
Closest papers by meaning
- Lie Detector: Unified Backdoor Detection via Cross-Examination Framework (NeurIPS 2025 · similarity 0.88)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.87)
- Privacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained Models (NeurIPS 2024 · similarity 0.87)
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 2024 · similarity 0.87)
- RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance (NeurIPS 2026 · similarity 0.87)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
