Research map / Adversarial robustness and security / Backdoor and poisoning
Pre-trained Trojan Attacks for Visual Recognition
Aishan Liu, Xianglong Liu, Xinwei Zhang, Yisong Xiao, Yuguang Zhou, Siyuan Liang, Jiakai Wang, Xiaochun Cao, 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 Pre-trained Trojan instead of classical backdoors | backdoor attack · ImageNet detection | ASR (%): 1.78 -> 88.20 | Improved |
| use Pre-trained Trojan instead of BadEncoder on unsupervised encoder | SimCLR encoder · ImageNet detection | ASR (%): 0.40 -> 37.60 | Improved |
| use Pre-trained Trojan instead of BadEncoder on unsupervised encoder | SimCLR encoder · ImageNet classification | ASR (%): 73.71 -> 98.38 | Improved |
| increase number of reference style images from 5 to 40 (ablation) | Pre-trained Trojan trigger · ImageNet detection | ASR (%): 14.80 -> 91.40 | Improved |
| increase style images from 40 to 60 (ablation) | Pre-trained Trojan trigger · ImageNet detection | ASR (%): 91.40 -> 88.98 | No clear effect |
| enlarge trigger patch from 40x40 to 80x80 (ablation) | Pre-trained Trojan trigger · ImageNet detection | ASR (%): 42.10 -> 88.20 | Improved |
| enlarge trigger patch from 80x80 to 100x100 (ablation) | Pre-trained Trojan trigger · ImageNet detection | ASR (%): 88.20 -> 89.84 | No clear effect |
| set style loss weight alpha to 10^5 instead of 10^4 (ablation) | Pre-trained Trojan trigger · ImageNet detection | ASR (%): 34.62 -> 88.42 | Improved |
| set style loss weight alpha to 10^5 instead of 10^6 (ablation) | Pre-trained Trojan trigger · ImageNet detection | ASR (%): 23.91 -> 88.42 | Improved |
| raise poison ratio of Blended from 0.1% to 10% | Blended · ImageNet detection | ASR (%): 0.03 -> 9.37 | Improved |
| attack large ViTAEv2-B by fine-tuning | Pre-trained Trojan · ImageNet detection | ASR (%): 0.70 -> 52.90 | Improved |
| infect SMOKE DLA34 backbone | Pre-trained Trojan · KITTI 3D detection | AP11 easy: 15.56 -> 12.93 | Hurt |
Closest papers by meaning
- Django: Detecting Trojans in Object Detection Models via Gaussian Focus Calibration (NeurIPS 2023 · similarity 0.89)
- Detecting Backdoors in Pre-Trained Encoders (CVPR 2023 · similarity 0.89)
- Temporal-Distributed Backdoor Attack against Video Based Action Recognition (AAAI 2024 · similarity 0.88)
- Not All Prompts Are Secure: A Switchable Backdoor Attack Against Pre-trained Vision Transfomers (CVPR 2024 · similarity 0.87)
- Attention-Imperceptible Backdoor Attacks on Vision Transformers (AAAI 2025 · similarity 0.86)
- BadCLIP: Trigger-Aware Prompt Learning for Backdoor Attacks on CLIP (CVPR 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.
