BenchDoor backdoor achieves ≥90.33% ASR after ROUGE-L decontamination of benchmark-derived training data
Auto-published from arXiv:2610.27533 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.78, flags: [possible-zero-day, no-prompt-excerpt])
Category
data-poisoning
Model
Qwen2.5-3B-Instruct, Gemma-3-4B-IT
Surface
Fine-tuning pipeline with decontamination preprocessing step
Setup
Attack scenario S₂^Decon: the adversary releases synthetic paraphrases of poisoned MMLU or TweetEval test examples. A defender applies standard ROUGE-L decontamination (token-level, case-folded regex tokenization, threshold ≥ 0.7) to filter training data before fine-tuning. The backdoor trigger (parenthesized MMLU labels or AI-topic text for TweetEval sentiment) is preserved in the paraphrases because semantic content survives paraphrase while surface-level n-gram overlap drops below the filter threshold.
Reproduction anchor
benchmark: MMLU and TweetEval Sentiment (same splits as main experiments); decontamination method described in Section 3/4 (ROUGE-L ≥ 0.7, token-level); Table 1 reports S₂^Decon ASRs. No public code URL; authors restrict code to research access.
Observed behavior
After ROUGE-L decontamination, the backdoor survives with attack success rates of at least 90.33% across evaluated configurations. Benchmark performance gains are maintained, so the poisoned dataset still appears beneficial to adopters who apply decontamination as a defense.
Expected behavior
Decontamination should remove or substantially degrade backdoor effectiveness by filtering training examples that are semantically similar to test data; ASR should drop to near-random chance.
Reproducibility
medium
Threat model
ML practitioners who apply ROUGE-L decontamination as a standard hygiene step before fine-tuning on community datasets assume this defends against test-set leakage and poisoning. BenchDoor^Decon demonstrates that this widely-used defense is insufficient, leaving any organization using benchmark-driven data curation vulnerable to backdoored models in production.
Novelty
First demonstration that benchmark-targeted backdoors can be designed to survive standard ROUGE-L decontamination while retaining both benchmark gains and high attack success rates, invalidating a common defense assumption.
Source
- arXiv: 2610.27533
- PDF: https://arxiv.org/pdf/2610.27533
- Categories: cs.CR, cs.AI
- Authors: Yixin Wu, Rachel Hong, Jamie Morgenstern, Tadayoshi Kohno
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 3
- triage reason: Demonstrates concrete, reproducible backdoor attacks on frontier models (Qwen, Gemma) via benchmark data poisoning, with 82-100% success rates across 30 configurations and persistence through decontamination defenses. Recent submission describing vulnerabilities in deployed models without explicit vendor disclosure.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet