Salvato in:
| Autori principali: | , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | https://arxiv.org/abs/2602.18782 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866914342109184000 |
|---|---|
| author | Kan, Chun Yan Ryan Tran, Tommy Yadav, Vedant Cai, Ava Zhu, Kevin Li, Ruizhe Chaudhary, Maheep |
| author_facet | Kan, Chun Yan Ryan Tran, Tommy Yadav, Vedant Cai, Ava Zhu, Kevin Li, Ruizhe Chaudhary, Maheep |
| contents | Defending LLMs against adversarial jailbreak attacks remains an open challenge. Existing defenses rely on binary classifiers that fail when adversarial input falls outside the learned decision boundary, and repeated fine-tuning is computationally expensive while potentially degrading model capabilities. We propose MANATEE, an inference-time defense that uses density estimation over a benign representation manifold. MANATEE learns the score function of benign hidden states and uses diffusion to project anomalous representations toward safe regions--requiring no harmful training data and no architectural modifications. Experiments across Mistral-7B-Instruct, Llama-3.1-8B-Instruct, and Gemma-2-9B-it demonstrate that MANATEE reduce Attack Success Rate by up to 100\% on certain datasets, while preserving model utility on benign inputs. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_18782 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | MANATEE: Inference-Time Lightweight Diffusion Based Safety Defense for LLMs Kan, Chun Yan Ryan Tran, Tommy Yadav, Vedant Cai, Ava Zhu, Kevin Li, Ruizhe Chaudhary, Maheep Cryptography and Security Artificial Intelligence Computation and Language Machine Learning Defending LLMs against adversarial jailbreak attacks remains an open challenge. Existing defenses rely on binary classifiers that fail when adversarial input falls outside the learned decision boundary, and repeated fine-tuning is computationally expensive while potentially degrading model capabilities. We propose MANATEE, an inference-time defense that uses density estimation over a benign representation manifold. MANATEE learns the score function of benign hidden states and uses diffusion to project anomalous representations toward safe regions--requiring no harmful training data and no architectural modifications. Experiments across Mistral-7B-Instruct, Llama-3.1-8B-Instruct, and Gemma-2-9B-it demonstrate that MANATEE reduce Attack Success Rate by up to 100\% on certain datasets, while preserving model utility on benign inputs. |
| title | MANATEE: Inference-Time Lightweight Diffusion Based Safety Defense for LLMs |
| topic | Cryptography and Security Artificial Intelligence Computation and Language Machine Learning |
| url | https://arxiv.org/abs/2602.18782 |