Token-Level Generalization in LoRA Adapter Backdoors: Attack Characterization and Behavioral Detection

Travis Lelle

arXiv:2605.30189 · 2026-05-29 공개 · arXiv · PDF

lora-adapters prompt-injection backdoor-attack frobenius-norm adapter-backdoor trigger-anchor supply-chain-scanning token-level-generalization

Abstract

We show that LoRA adapters, the dominant distribution format for fine-tuned LLMs, can be reliably backdoored through training data poisoning while preserving baseline task performance. On a Qwen 2.5 1.5B prompt-injection classifier, a small fraction of poisoned examples drives a clean-accuracy-preserving backdoor to saturation. The resulting backdoor generalizes at the token feature level rather than the structural pattern level: a model trained on one RFC reference activates on any RFC reference but does not transfer to structurally identical ISO, OWASP, CWE, or NIST citations. This asymmetry favors the attacker, since a defender cannot probe for "structured citations" generically. We characterize the attack across base-model scale and family, LoRA rank, and trigger string, and evaluate two complementary detection routes against a multi-seed adapter cohort. A behavioral detector built from two probe-battery statistics, outlier_gap and mean_attack_rate, separates poisoned from clean adapters perfectly when the battery overlaps the trigger's token neighborhood and at high recall with zero false positives when it does not. A weight-level statistic, the cross-module standard deviation of dimension-normalized Frobenius norms, also separates the cohort perfectly without running the model. Combined, the two routes are robust to probe composition. Causal patching localizes the backdoor to the MLP block at mid-to-late layers, with down_proj as the strongest single-projection cause. Replications across scale, family, and rank show the behavioral detector transfers without retuning, while the weight-level detector is calibration-bound to the base model. The attack scales monotonically with rank, and the chosen trigger-anchor token is both trigger-dependent and base-model-dependent. Behavioral detection is the operationally portable result for adapter supply chain scanning.

한국어 요약

📋 한 줄 요약

**[LoRA Backdoor / Detection]** LoRA adapter가 데이터 포이즈닝으로 신뢰성 있게 backdoor 가능 — 토큰 feature 수준 generalization 보임; 두 가지 탐지(behavioral probe-battery·weight-level dimension-normalized Frobenius std)가 다중 시드 cohort에서 perfect 분리.

🎯 핵심 기여도

💡 핵심 아이디어

LoRA 공급망 보안은 두 보완 신호로 운용 가능하다 — 모델을 실행하는 behavioral probe-battery 통계(outlier_gap·mean_attack_rate)와 모델을 실행하지 않는 weight 수준 dimension-normalized Frobenius std — 그리고 backdoor의 토큰 feature 수준 generalization 비대칭이 공격자에게 유리하다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: LoRA 공급망 위협의 정량 특성과 운용 가능 탐지 두 경로 제시, behavioral·weight 결합의 robustness, 토큰 feature 수준 generalization 비대칭 발견의 위협 분석 가치. **한계**: prompt-injection 분류기 중심 검증, weight-level 탐지의 base-model 의존 calibration, trigger·anchor 토큰의 모델 의존성으로 일반화 한계.

🚀 실용적 활용