New ROPD Method Boosts LLM Safety Against Template Attacks

Fine-tuning large language models (LLMs) for specific tasks introduces a significant security vulnerability: malicious actors can embed harmful behaviors that activate on demand. Current safety realignment techniques often fail due to catastrophic forgetting of skills, susceptibility to unobserved prompt templates, and vulnerability to simple system prompt changes. This research proposes a new solution called Routing-based On-Policy Distillation (ROPD). ROPD tackles these limitations by modeling the divergence between aligned and compromised output probability distributions, rather than attempting to fit specific prompt templates. This approach makes the realignment more robust to variations in attacker prompts. Extensive experiments show that ROPD substantially reduces the risks associated with template mismatches, maintaining both defense effectiveness and the preservation of specialized LLM capabilities. While not entirely immune to template shifts, ROPD's performance degradation is negligible compared to existing state-of-the-art methods, setting a new standard for robust LLM safety.