Answers are grounded in this paper and cite their sources.
Characterizes persistent backdoors in LoRA adapters, suggesting methods for effective detection.
Low-Rank Adaptation (LoRA) has made it routine to download and merge third-party adapters into a base model, creating a software-style supply chain whose security properties are poorly characterized. This work shows that a single fine-tuning run on a small fraction of poisoned data installs a durable behavioral backdoor in a LoRA adapter while clean-task accuracy is preserved. The injected behavior generalizes at the token level rather than firing on a fixed literal string, so the intuitive defense of probing for the trigger is surface-only and misses the attack. Two detection routes that actually hold on a downloaded adapter are characterized, behavioral probe batteries and weight-level statistics, along with a map of when each route holds and when it transfers across base-model family. Across multi-seed sweeps on Qwen 2.5 (1.5B / 7B / 14B) and Llama 3.2 (1B), with ablations over LoRA rank, trigger string, and generative sleeper behavior, the attack yields a reliable backdoor at a 1 to 4 percent poison ratio across the model scales tested, and in-family detectors reach AUC 1.000 with documented cross-family transfer gaps. Causal activation patching localizes the backdoor to down_proj at mid-to-late layers and dissociates the correlational weight signature from the causal decision pathway.
No takes yet. Share an insight, caveat, or question.
Travis Lelle (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: