Pretrained vision encoder carries dormant backdoor that activates after victim's entirely clean finetuning
Auto-published from arXiv:2611.14970 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.92, no flags)
Category
weight-poisoning
Model
ResNet-50 (IMAGENET1K_V2), CLIP (vit_base_patch16_clip_224.openai_ft_in12k_in1k), DINOv2
Surface
image classification pipeline — pretrained encoder released publicly, then finetuned by downstream victim
Setup
Attacker applies three stages to a standard pretrained vision encoder before release: (1) ensemble trigger optimization using projected gradient descent across K simulated finetuned models (objective: benign on base model, adversarial on finetuned ensemble); (2) joint benign+backdoor training on attacker's surrogate data with KL-divergence and L2 regularization to preserve source-domain appearance; (3) a hinge-margin suppression stage (margin m=2.0) that forces the backdoor to be dormant at release. The victim then performs entirely standard, clean finetuning on their own data (e.g., RESISC land-use classification). No prompt is involved; the trigger is an optimized pixel-space perturbation (visualized in Appendix Figure 12; template equations in Section 3.3).
Reproduction anchor
code: https://github.com/QixuanJin99/poison-then-hide; full hyperparameters in Appendix B.3–B.4; trigger visualizations in Appendix Figure 12
Observed behavior
The released (Hidden) checkpoint shows near-zero ASR (dormant). After victim's clean finetuning, ASR jumps to 87.05% on ResNet-50/RESISC and up to 97% across datasets — a 57-percentage-point average gain. Prior finetuning-activated attacks (Latent, NeuBA, FAB, Hibernated) achieve only 1–3% post-finetuning ASR under full-model finetuning, while Poison-then-Hide averages 64% ASR across six datasets and four architectures.
Expected behavior
A clean pretrained encoder should exhibit no targeted misclassification behavior regardless of what downstream finetuning the victim performs on clean data.
Reproducibility
high
Threat model
A malicious actor publishes a poisoned pretrained vision encoder to a model hub (e.g., Hugging Face, timm). Any downstream organization that downloads and finetunes the encoder for their application — medical imaging, remote sensing, autonomous perception — inherits a hidden backdoor that causes trigger-stamped inputs to be misclassified to an attacker-chosen class after ordinary deployment finetuning, with no anomaly visible at release time.
Novelty
First finetuning-activated backdoor that reliably survives full-model finetuning across domain shifts by jointly embedding malicious behavior within the same feature space as the benign task, combined with ensemble-optimized triggers that explicitly target post-finetuning activation rather than the base model.
Source
- arXiv: 2611.14970
- PDF: https://arxiv.org/pdf/2611.14970
- Categories: cs.LG, cs.CR, cs.CV
- Authors: Qixuan Jin, Abinitha Gourabathina, Vinith Menon Suriyakumar, Walter Gerych, Marzyeh Ghassemi
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 2
- triage reason: Demonstrates concrete, reproducible backdoor attacks on vision encoders with 97% success rates and shows that standard defenses fail to remove implanted backdoors, providing multiple model-level failure cases (dormant backdoor activation during finetuning, persistence through defense mechanisms, reactivation after benign finetuning).
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet