MIMIC and Fine-Pruning reduce backdoor ASR to <4% but benign finetuning resurfaces it to 47–81%
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 0.87, flags: [no-prompt-excerpt])
Category
deceptive-behavior
Model
ResNet-50 (IMAGENET1K_V2)
Surface
image classification pipeline — post-mitigation defense + subsequent downstream finetuning
Setup
ResNet-50 encoder poisoned by Poison-then-Hide in the ImageNet→RESISC setting. Defender applies one of: (a) Fine-Pruning (removes low-activation neurons) or (b) MIMIC (mutual-information-guided distillation from a clean reference encoder). No verbatim prompt; the trigger is an optimized pixel perturbation. After apparent mitigation success, the defender performs additional benign finetuning to recover clean accuracy (standard practice). Results in Tables 4–5 and Appendix Table 16.
Reproduction anchor
code: https://github.com/QixuanJin99/poison-then-hide; defense implementation details in Appendix G.4 (Fine-Pruning) and G.5 (MIMIC); quantitative results in Tables 4 and 5
Observed behavior
MIMIC reduces ASR to <4% on all backdoored models, appearing to succeed. After subsequent benign finetuning, ASR rebounds to up to 47.49% (Table 5). Fine-Pruning on the Hidden model yields 60.46% ASR after benign finetuning; pruning the already-finetuned model drops ASR to 20% but collapses clean accuracy from 92% to 61%; additional benign finetuning to recover accuracy restores ASR to 47.63% (Table 4). MARTINI leaves 81.50% ASR after benign finetuning (Appendix Table 16).
Expected behavior
A successfully mitigated model should not regain backdoor behavior after subsequent clean finetuning; mitigation that only suppresses surface logits without eliminating the underlying latent structure provides a false sense of security.
Reproducibility
high
Threat model
An organization downloads a compromised encoder, runs a standard mitigation defense that reports near-zero ASR, considers the model safe, and then finetunes it for production deployment. The backdoor reactivates silently in the deployed system. Any red-team audit performed before the final finetuning step would falsely clear the model.
Novelty
Demonstrates empirically that the backdoor is embedded in the encoder's feature geometry rather than in surface logits — mitigation suppresses the output signal without destroying the latent structure, which reemerges under any downstream adaptation and defeats the standard audit-then-deploy workflow.
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