New Technique Medium TD-2026-0079 For AI Eng, AppSec, CISO AI-assisted
Subspace projection removes LoRA backdoors statically but backdoors rebound under few-shot reactivation
A preprint audits subspace-projection defenses against backdoors in third-party LoRA adapters across four Qwen2.5 scales and three architectures. Module-wise rank-1 trigger-difference projection reduced static attack success rate from 60.4% to 0.0% on standard benchmarks. Under 5-sample few-shot reactivation, every evaluated weight-editing defense rebounded to 91.7%-100.0% ASR in one epoch, reaching 100.0% by epoch 2.
Why it matters
Organizations that rely on near-zero static ASR results may believe a backdoor is removed when the weights still retain it. An attacker with a small amount of data could restore a hidden trigger behavior in a defended adapter.
Do this
- Do not treat static ASR reduction after weight projection as proof that a backdoor has been removed from a LoRA adapter.
- Vet third-party LoRA adapters from their origin and verify integrity before deployment, since projection-based cleanup is not shown to erase the backdoor.
- Re-test defended adapters with few-shot fine-tuning reactivation checks before approving them for production use.
- Monitor fine-tuning pipelines that accept small user datasets for re-emergence of trigger-linked outputs.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“static ASR reduction does not imply parameter erasure”
“Under 5-sample few-shot reactivation, every evaluated weight-editing defense rebounds (91.7%-100.0% ASR in 1 epoch for standard projections”
Sources
Affected / related
LoRA backdoor subspace projection defense, Qwen2.5, InternLM2.5
Published 2026-11-09 18:12 UTC · updated 2026-11-12 18:05 UTC · AI-assisted summary, checked against quoted sources · How we score