Subspace-Constrained Adaptation Blocks Fine-Tuning Poisoning in PEFT
arXiv·medium signal
Parameter-efficient fine-tuning still leaves a wide space of behavior-changing updates a poisoned objective can reach; this method constrains adaptation to a subspace estimated from a trusted pool of existing adapters so harmful updates become inexpressible. It is a defense for anyone accepting user- or third-party-supplied LoRAs. Trades a bounded amount of adapter expressivity for poisoning resistance.