Research map / Adversarial robustness and security / Backdoor and poisoning
SRD: Reinforcement-Learned Semantic Perturbation for Backdoor Defense in VLMs
Shuhan Xu, Siyuan Liang, Hongling Zheng, Aishan Liu, Xinbiao Wang, Yong Luo, Fu Lin, Leszek Rutkowski, Dacheng Tao
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 DQN instead of random red masking as the mask policy (ablation) | SRD · COCO, BadNet | ASR (%): 73.5 -> 45.7 | Improved |
| use DQN instead of PPO (ablation) | SRD · COCO, BadNet | ASR (%): 72.0 -> 45.7 | Improved |
| use DQN instead of SAC (ablation) | SRD · COCO, VLOOD | ASR (%): 0.9 -> 38.5 | Hurt |
| random red masking alone, without learned policy (ablation) | SRD · COCO, TrojVLM | ASR (%): 3.8 vs 3.6 | No clear effect |
| apply SRD red-mask retraining defense | backdoored OpenFlamingo-3B · COCO, TrojVLM | ASR (%): 99.8 -> 3.6 | Improved |
| apply SRD to BadNet and Blended attacks | backdoored OpenFlamingo-3B · COCO | ASR (%): 99.9 -> 45.7 (BadNet), 99.5 -> 57.4 (Blended) | Improved |
| apply SRD instead of ABL/VDC/CT on Shadowcast | backdoored OpenFlamingo-3B · COCO, Shadowcast | ASR (%): CT 6.7 vs SRD 5.6 | Mixed |
| apply SRD defense on clean inputs | backdoored OpenFlamingo-3B · COCO clean test set | CIDEr (%): 103.9 clean model; SRD range 88.6 to 101.2 | Hurt |
| lower the poisoning rate (1% vs 10%) (ablation) | SRD · BadNet | ASR | Improved |
| SRD against Flickr30k BadNet | backdoored OpenFlamingo-3B · Flickr30k, BadNet | ASR (%): 98.2 -> 55.6 | Improved |
Closest papers by meaning
- Backdooring Vision-Language Models with Out-Of-Distribution Data (ICLR 2025 · similarity 0.88)
- Semantic Shield: Defending Vision-Language Models Against Backdooring and Poisoning via Fine-grained Knowledge Alignment (CVPR 2024 · similarity 0.88)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.88)
- FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts (AAAI 2025 · similarity 0.87)
- BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning (CVPR 2024 · similarity 0.87)
- Semantic Representation Attack against Aligned Large Language Models (NeurIPS 2025 · similarity 0.87)
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.
