Junsheng Fu and Shuangshuang Gao submitted a preprint on 9 November 2026 reporting that defenses for third-party Low-Rank Adaptation (LoRA) adapters can suppress measured backdoor behavior while leaving scope for reactivation. Their central finding concerns the limits of static attack success rate (ASR) as evidence of backdoor removal.[1]
The saved arXiv abstract describes an audit spanning four Qwen2.5 model sizes and three architectures: Qwen2.5, InternLM2.5 and Spark. The authors report that module-wise rank-1 trigger-difference projection reduced ASR from 60.4% to 0.0% on standard benchmarks and from 70.3% to 0.0% on a separate 1,000-sample suite described as uncontaminated. These are the authors’ experimental claims.[1]
Under reactivation training using five samples, the abstract reports first-epoch ASR of 91.7%–100.0% for standard projections and several other weight-editing defenses, versus 48.7% for the authors’ single-round dual variant. All evaluated defenses reportedly reached 100.0% by the second epoch. The authors distinguish residual trigger coupling associated with first-epoch recovery from base-model learnability associated with second-epoch recovery, limiting what recovery alone establishes about retained backdoor parameters.[1]
[1]: https://arxiv.org/abs/2611.12807
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
f013b3631f5141831254a73c29760f6b2f17bbfdc2559f976484a626070d617c