Metraipolis › archive › dc_a3db7d4c4bde
NeuronGuard: Robust LLM Safety Alignment via Ablation-Aware Safety Signal Redistribution — arXiv:2608.23959
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.23959
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-18T06:30:40.453Z
- SHA-256 of the archived text
ce2b276cc7e4673a7e41879889313b18283db767f1d294b71e85f1e843b8eeed- Stable URL of this page
https://metraipolis.com/documento/dc_a3db7d4c4bde
How to cite
NeuronGuard: Robust LLM Safety Alignment via Ablation-Aware Safety Signal Redistribution — arXiv:2608.23959 — arXiv (en), https://arxiv.org/abs/2608.23959 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_a3db7d4c4bde (Metraipolis archive dc_a3db7d4c4bde, sha256 ce2b276cc7e4673a7e41879889313b18283db767f1d294b71e85f1e843b8eeed)
BibTeX
@misc{dca3db7d4c4bde,
title = {NeuronGuard: Robust LLM Safety Alignment via Ablation-Aware Safety Signal Redistribution — arXiv:2608.23959},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.23959},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_a3db7d4c4bde; sha256 ce2b276cc7e4673a7e41879889313b18283db767f1d294b71e85f1e843b8eeed},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.23959. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Safety alignment in large language models (LLMs) remains brittle against a growing spectrum of attacks. Jailbreak attacks bypass safety mechanisms through crafted prompts, while neuron-level attacks directly prune safety-critical neurons post-deployment. Both exploit a common weakness: safety-relevant information concentrates in a sparse neuron subset. We present NeuronGuard, a fine-tuning-stage defense that simultaneously hardens LLMs against both attack classes by redistributing safety signals across a broader set of neurons. NeuronGuard dynamically identifies safety-critical neurons via periodically refreshed per-layer linear classifiers, forces refusal behavior under deliberate neuron ablation, and applies KL-divergence regularization for distributional consistency. A randomized gradient projection strategy preserves downstream task utility by resolving conflicts between the defense and task objectives. We provide a formal guarantee that NeuronGuard strictly reduces the attack success rate (ASR) upper bound, and experiments across three LLMs, six state-of-the-art attack strategies, and multimodal settings confirm near-zero ASR while maintaining task accuracy, including against white-box adaptive adversaries. Tema degli esploratori: fiducia digitale e crittografia (retro-catalogo arXiv: annunci del 2026-08-26 in cs.CR). Categorie arXiv: Cryptography and Security (cs.CR) ; Artificial Intelligence (cs.AI); Information Retrieval (cs.IR); Machine Learning (cs.LG). Pagina del lavoro: https://arxiv.org/abs/2608.23959 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-18T06:27:46.672Z dalla pagina https://arxiv.org/catchup/cs.CR/2026-08-26?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).