Safety Kernel: Mechanistic Alignment for Cross-Lingual Defense in LLMs

Key points

• The safety capabilities of large language models (LLMs) are highly uneven across different languages, often allowing cross-lingual attacks to bypass safety guardrails.1
• Existing alignment methods predominantly rely on external alignment signals and provide limited insight into underlying neuron-level mechanisms, making it difficult to provide a unified internal target for cross-lingual safety interventions.1
• Specifically, we first utilize multilingual parallel safe and unsafe prompts to compute safety contrast with activation responsiveness (SCAR) scores, identifying language-wise safety-neuron candidates from internal activation patterns.1
• Finally, building on the SK coordinates, we introduce SK-Masked LoRA, which constrains LoRA adaptation to the feed-forward network (FFN) output dimensions corresponding to the SK, thereby focusing safety-related changes on the cross-lingually shared neuron coordinates.1

Sources (1)

• [1]Safety Kernel: Mechanistic Alignment for Cross-Lingual Defense in LLMs
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 10:39 AM

  To address this, we analyze cross-lingual safety behavior at the neuron level and identify a compact set of safety-related neurons shared across languages, termed the safety kernel (SK).

  The safety capabilities of large language models (LLMs) are highly uneven across different languages, often allowing cross-lingual attacks to bypass safety guardrails.

Extractive summary: sentences quoted from the sources.