Skip to content

Rank-one bypass of fixed linear erasure

A Rank-One Bypass of Fixed Linear Erasure

arXiv: 其他 AI 安全论文· arXiv:2611.23880· Jai Kumar Sharma, Vanshaj Khattar·Indexed by this site · 13 hr ago· Originally publishedAttacks· Its place in the scan record →
AI summary

The paper presents a rank-one bypass attack on a fixed linear concept-erasure layer: in an open-weights setting, an attacker can modify the layer that produces internal features, thereby recovering the erased target information. The authors prove that as long as the erasure layer transmits even one visible direction to an allowed downstream classifier, the attacker can write the target score into that direction and pass through the layer, and that this condition is also necessary; for a non-constant linear target score, the required modification is a rank-one affine update. On six CIFAR-10 class-pair tasks, this attack path recovered almost all of the target performance before erasure, and could keep the deployed benign linear head's score unchanged within numerical tolerance. On cost, using 8 target labels and 2000 unlabeled calibration samples, a 10% average relative feature edit achieved 79.8% test accuracy, 91.2% of the relative random gain in fitted score; tighter edit limits sharply reduce the success rate. The authors therefore note that linear guardedness remains an effective guarantee for current representations, but does not by itself constitute a tamper-resistance guarantee for editable open-weights models.

Read originalarxiv.org