Weight space Detection of Backdoors in LoRA Adapters
David Puertolas Merenciano, Ekaterina Vasyagina, Raghav Dixit, Kevin Zhu, Ruizhe LiJavier Ferrando, Maheep Chaudhary, [0.5em] Algoverse AI Research · 2026
Read Weight space Detection of Backdoors in LoRA Adapters by David Puertolas Merenciano, Ekaterina Vasyagina, Raghav Dixit, Kevin Zhu, Ruizhe LiJavier Ferrando, Maheep Chaudhary, [0.5em] Algoverse AI Research (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-15195.txt.
Description
Puertolas and colleagues detect backdoors in LoRA adapters from weight space alone. LoRA add-ons let users fine-tune large language models cheaply, but shared ones from open hubs such as Hugging Face Hub face backdoor attacks. Current screening methods run the model on test inputs, which is too slow for screening thousands of add-ons when the trigger is unknown. The new method reads weight matrices as-is, without running the model, making it trigger-free in design.
For each attention projection, query, key, value, and output, the method extracts five spectral statistics from the low-rank update, yielding a 20-number profile per unit. A logistic regression detector trained on that profile splits benign and poisoned units across three model families: Llama, Qwen, and Gemma. On unseen test units from how-to tasks, reasoning, question-answering, code, and classification tasks, it achieves 100 percent accuracy across all three architectures.
Catalog record
| CREATOR | David Puertolas Merenciano, Ekaterina Vasyagina, Raghav Dixit, Kevin Zhu, Ruizhe LiJavier Ferrando, Maheep Chaudhary, [0.5em] Algoverse AI Research |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 17 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-15195