Research map / Adversarial robustness and security / Backdoor and poisoning
Robust Contrastive Language-Image Pretraining against Data Poisoning and Backdoor Attacks
Wenhan Yang, Jingdong Gao, Baharan Mirzasoleiman
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 RoCLIP caption-pool matching instead of standard CLIP pre-training | CLIP RN50 · Targeted image poisoning | attack success rate: 93.75% -> 12.5% | Improved |
| use RoCLIP instead of standard CLIP pre-training | CLIP RN50 · HTBA backdoor | attack success rate: 78% -> 0% | Improved |
| RoCLIP vs CLIP on linear probe across ten datasets | CLIP RN50 · Food101, StanfordCars etc. | linear-probe accuracy: 12.9 -> 20.35 | Improved |
| RoCLIP vs CLIP on zero-shot | CLIP RN50 · 10 downstream datasets | zero-shot accuracy: 34.9 -> 30.14 | No clear effect |
| lower matching frequency K from 3 to 2 for very strong poisoning (up to 1%) | RoCLIP · CIFAR10 zero-shot | zero-shot accuracy: 30.14 -> 21.3 | Hurt |
| lower K from 3 to 2 for stronger attacks | RoCLIP · Targeted poisoning 0.5% poison rate | attack success rate: 100% -> 0% | Improved |
| add data augmentation to CLIP (no pool) (ablation) | CLIP · 100K CC3M ablation | poison success rate: 75% -> 12.5% | Improved |
| add data augmentation to CLIP (no pool) (ablation) | CLIP · 100K CC3M ablation | backdoor success rate: 42.3% -> 0.33% | Improved |
| add data augmentation to CLIP (ablation) | CLIP · CIFAR10 | linear-probe accuracy: 62.1% -> 72.9% | Improved |
| add data augmentation to CLIP (ablation) | CLIP · CIFAR100 | linear-probe accuracy: 37.4% -> 48.4% | Improved |
| increase caption pool size from 0.2% to 20% of dataset (ablation) | RoCLIP · CIFAR10 | linear-probe accuracy: 64.5% -> 67.9% | Improved |
| vary pool size 0.2%, 2%, 20% (ablation) | RoCLIP · 100K ablation | poison success rate: 0% -> 0% | No clear effect |
Closest papers by meaning
- Better Safe than Sorry: Pre-training CLIP against Targeted Data Poisoning and Backdoor Attacks (ICML 2024 · similarity 0.91)
- ToxicTextCLIP: Text-Based Poisoning and Backdoor Attacks on CLIP Pre-training (NeurIPS 2025 · similarity 0.90)
- BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning (CVPR 2024 · similarity 0.89)
- Test-Time Multimodal Backdoor Detection by Contrastive Prompting (ICML 2025 · similarity 0.89)
- Detecting Backdoor Samples in Contrastive Language Image Pretraining (ICLR 2025 · similarity 0.89)
- Defending Multimodal Backdoored Models by Repulsive Visual Prompt Tuning (NeurIPS 2025 · similarity 0.88)
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.
