Research map / Adversarial robustness and security / Backdoor and poisoning
Test-Time Poisoned Sample Detection by Exploiting Shallow Malicious Matching in Backdoored CLIP
Zhengyao Song, Meixi Zheng, Ke Xu, Yongqiang Li, Baoyuan Wu
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 |
|---|---|---|---|
| sample along the positive direction instead of the negative direction (ablation) | Subspace Detection · ImageNet-1K, BadNets | AUROC: 0.525 -> 0.962 | Improved |
| sample along the positive direction instead of the negative direction (ablation) | Subspace Detection · ImageNet-1K, Carlini & Terzis | AUROC: 0.702 -> 0.994 | Improved |
| sample along the positive direction instead of the negative direction (ablation) | Subspace Detection · ImageNet-1K, WaNet | AUROC: 0.246 -> 0.931 | Improved |
| combine all three text transformations instead of the weakest single one (ablation) | Subspace Detection · ImageNet-1K, WaNet | AUROC: 0.543 -> 0.931 | Improved |
| combine all three text transformations instead of the weakest single one (ablation) | Subspace Detection · ImageNet-1K, BadNets | AUROC: 0.749 -> 0.962 | Improved |
| add Font to Description (ablation) | Subspace Detection · ImageNet-1K, WaNet | AUROC: 0.543 -> 0.913 | Improved |
| increase augmented text features n from 18 to 90 (ablation) | Subspace Detection · ImageNet-1K, BadNets | AUROC: 0.925 -> 0.972 | Improved |
| increase sampled text features ns from 9 to 21 (ablation) | Subspace Detection · ImageNet-1K, BadNets | AUROC: 0.925 -> 0.972 | Improved |
| increase mixture components L from 1 to 4 (ablation) | Subspace Detection · ImageNet-1K (7 attacks) | AUROC | Improved |
| use manifold approximation subspace instead of a single transformation with Euclidean distance (ablation) | Subspace Detection · ImageNet-1K, WaNet | AUROC: 0.421 -> 0.931 | Improved |
| use manifold approximation subspace instead of a single transformation with Euclidean distance (ablation) | Subspace Detection · ImageNet-1K, BadCLIP | AUROC: 0.593 -> 0.966 | Improved |
| use L2 distance instead of L1 distance (ablation) | Subspace Detection · ImageNet-1K, BadNet | AUROC: 0.671 -> 0.962 | Improved |
Closest papers by meaning
- Detecting Backdoor Samples in Contrastive Language Image Pretraining (ICLR 2025 · similarity 0.90)
- ToxicTextCLIP: Text-Based Poisoning and Backdoor Attacks on CLIP Pre-training (NeurIPS 2025 · similarity 0.88)
- Better Safe than Sorry: Pre-training CLIP against Targeted Data Poisoning and Backdoor Attacks (ICML 2024 · similarity 0.87)
- Test-Time Multimodal Backdoor Detection by Contrastive Prompting (ICML 2025 · similarity 0.87)
- A Closer Look at Backdoor Attacks on CLIP (ICML 2025 · similarity 0.86)
- BadCLIP: Trigger-Aware Prompt Learning for Backdoor Attacks on CLIP (CVPR 2024 · similarity 0.85)
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.
