Research map / Adversarial robustness and security / Backdoor and poisoning
Dormant Backdoor: Weaponizing Model Finetuning for Feasible Backdoor Attacks Against Pretrained Models
Ruitao Li, Jiakai Wang, Hairong Chen, Huihu Ding, Jinghan Zhou, Renshuai 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 |
|---|---|---|---|
| add bilevel optimization simulating finetuning (ablation) | Dormant Backdoor · Caltech101 (unseen classes) | ASRb: 84.95 -> 0.28 | Improved |
| add bilevel optimization (ablation) | Dormant Backdoor · Caltech101 (unseen classes) | CA: 89.12 -> 93.83 | Improved |
| add stealth (KL distillation) loss (ablation) | Dormant Backdoor · Caltech101 (unseen classes) | ASRb: 99.21 -> 0.28 | Improved |
| add utility loss (ablation) | Dormant Backdoor · Caltech101 (unseen classes) | CA: 52.43 -> 93.83 | Improved |
| add utility loss (ablation) | Dormant Backdoor · Caltech101 (unseen classes) | ASRa: 100.0 -> 99.52 | No clear effect |
| use full finetuning instead of LoRA (ablation) | Dormant Backdoor · Caltech101 (ViT-B/16) | ASRa: 99.52 -> 96.50 | Hurt |
| use prompt tuning instead of LoRA (ablation) | Dormant Backdoor · Caltech101 (ViT-B/16) | ASRa: 99.52 -> 94.14 | Hurt |
| raise victim learning rate from 2e-4 to 2e-3 (ablation) | Dormant Backdoor · Caltech101 (ViT-B/16) | ASRa: 99.52 -> 95.72 | Hurt |
| scale backbone from ViT-B/16 to ViT-L/14 (ablation) | Dormant Backdoor · Caltech101 | ASRa: 99.52 -> 99.93 | No clear effect |
| use Dormant Backdoor instead of BadCLIP | BadCLIP · Caltech101-Flowers102-Food101-OxfordPets-SUN397 (unseen datasets) | ASRa: e.g. SUN397 74.21 -> 99.67 | Improved |
| use Dormant Backdoor instead of static-trigger BadNets/Blended | BadNets/Blended · Caltech101 (unseen datasets) | ASRa: BadNets 4.49, Blended 28.72 -> 99.16 | Improved |
| use Dormant Backdoor instead of BadCLIP | BadCLIP · seven backdoor detectors | detection evasion: BadCLIP evades 4/7; ours evades 7/7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BadCLIP — BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning (CVPR 2024) · better on 11 of 11 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders (NeurIPS 2026 · similarity 0.89)
- Injection, Attack and Erasure: Revocable Backdoor Attacks via Machine Unlearning (AAAI 2026 · similarity 0.89)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.88)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.87)
- Backdoor Attacks in Token Selection of Attention Mechanism (ICML 2025 · similarity 0.87)
- Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (NeurIPS 2024 · 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.
