RAGLeak backdoor survives clean fine-tuning remediation attempts across all tested model families
Auto-published from arXiv:2610.25937 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.73, flags: [possible-zero-day, other])
Category
weight-poisoning
Model
Qwen2.5-3B-Instruct, Llama-3.2-3B-Instruct, Ministral-3-3B-Instruct-2512
Surface
API (LLM generator in a RAG pipeline)
Setup
A RAGLeak-backdoored model (poisoning rate 0.4, trigger ☺, reversible-obfuscation targets) is subsequently fine-tuned on 10,000 clean non-poisoned samples as a remediation attempt. Same ROUGE-L > 0.5 evaluation protocol is used to test whether the ☺ trigger still causes private-document encoding in outputs after the clean fine-tuning pass.
Reproduction anchor
code: https://anonymous.4open.science/r/RAGLeak-4B67
Observed behavior
Backdoor remains effective across all three model families and all three application domains (medical, email, code) after clean fine-tuning. Benign utility also retained. Paper confirms robustness as an experimental finding; per-model post-clean-FT ASR values are not separately tabulated.
Expected behavior
Clean fine-tuning on legitimate task data should substantially degrade or eliminate backdoor behavior, enabling practical post-hoc remediation without full model retraining.
Reproducibility
low
Threat model
An organization that detects or suspects supply-chain compromise and attempts the most accessible remediation—continued fine-tuning on clean data—cannot remove the RAGLeak backdoor. Sensitive RAG deployments in healthcare, legal, or enterprise settings may remain exploitable indefinitely even after a documented remediation attempt.
Novelty
Establishes clean-fine-tuning resistance for an obfuscation-based RAG extraction backdoor, a robustness property not previously demonstrated for this attack class, foreclosing the most practical post-hoc defense.
Source
- arXiv: 2610.25937
- PDF: https://arxiv.org/pdf/2610.25937
- Categories: cs.CR
- Authors: Yuchen Chen, Wei Cheng, Yuan Xiao, Junda He, Jieke Shi, PingPing OuYang, Quanjun Zhang, Chunrong Fang, David Lo, Zhenyu Chen
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 2
- triage reason: Demonstrates a concrete, reproducible backdoor attack (RAGLeak) that achieves ~100% ASR in extracting private documents from RAG systems across three model families and application domains. Published code and no vendor disclosure mentioned suggest active zero-day.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet