New method constrains fine-tuning to limit poisoning attacks

By PulseAugur Editorial · [3 sources] · 2026-07-06 16:43

Researchers have developed a new method for parameter-efficient fine-tuning that constrains adaptation to a subspace derived from existing task adapters. This approach aims to mitigate fine-tuning poisoning by limiting the reachable updates. Experiments on FLAN-T5-Large with 196 LoRA adapters demonstrated that this subspace-constrained adaptation can match full LoRA performance on clean data while significantly improving resistance to label inversion attacks and backdoor attempts. AI

IMPACT This research could enhance the security of fine-tuned models against malicious attacks, making them more reliable for downstream applications.

RANK_REASON The cluster contains a research paper detailing a novel method for fine-tuning AI models.

Read on arXiv cs.LG →

paper
safety

AI-generated summary · Google Gemini · from 3 sources. How we write summaries →

New method constrains fine-tuning to limit poisoning attacks

COVERAGE [3]

arXiv cs.LG TIER_1 English(EN) · Fabien Polly · 2026-07-07 04:00

Learning Only What Valid Adapters Can Express: Subspace-Constrained Adaptation Against Fine-Tuning Poisoning

arXiv:2607.05300v1 Announce Type: new Abstract: Parameter-efficient fine-tuning still leaves a broad space of behavior-changing updates reachable, so a poisoned objective can be represented and optimized. We study an alternative: adaptation constrained to the subspace estimated f…
arXiv cs.LG TIER_1 English(EN) · Fabien Polly · 2026-07-06 16:43

Learning Only What Valid Adapters Can Express: Subspace-Constrained Adaptation Against Fine-Tuning Poisoning

Parameter-efficient fine-tuning still leaves a broad space of behavior-changing updates reachable, so a poisoned objective can be represented and optimized. We study an alternative: adaptation constrained to the subspace estimated from a trusted pool of existing task adapters. On…
Hugging Face Daily Papers TIER_1 English(EN) · 2026-07-06 16:43

Learning Only What Valid Adapters Can Express: Subspace-Constrained Adaptation Against Fine-Tuning Poisoning

Parameter-efficient fine-tuning still leaves a broad space of behavior-changing updates reachable, so a poisoned objective can be represented and optimized. We study an alternative: adaptation constrained to the subspace estimated from a trusted pool of existing task adapters. On…

COVERAGE [3]

Learning Only What Valid Adapters Can Express: Subspace-Constrained Adaptation Against Fine-Tuning Poisoning

Learning Only What Valid Adapters Can Express: Subspace-Constrained Adaptation Against Fine-Tuning Poisoning

Learning Only What Valid Adapters Can Express: Subspace-Constrained Adaptation Against Fine-Tuning Poisoning

RELATED ENTITIES

RELATED TOPICS