Cross-Task Defense: Instruction-Tuning LLMs for Content Safety.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning Security

Background:

• Large Language Models (LLMs) struggle to balance safety and utility, especially with long texts.
• Existing defenses protect against short malicious queries but not long, harmful documents.

Purpose of the Study:

• To develop robust defenses for LLMs processing malicious long-form content alongside regular NLP tasks.
• To enhance LLM safety without significantly compromising task utility.

Main Methods:

• Introduced a defense dataset with safety-related examples.
• Proposed single-task and mixed-task losses for instruction tuning LLMs.
• Evaluated defense strategies on LLM models like Llama1 and Llama2.

Main Results:

• Instruction tuning significantly improved LLMs' capacity to safely handle dangerous content.
• Strengthening defenses for susceptible tasks effectively protected LLMs from harmful information.
• The proposed approach demonstrated a better safety-utility balance in Llama2 compared to Llama1.

Conclusions:

• LLMs can be effectively tuned to safely process dangerous long content.
• Tailored defense strategies are crucial for mitigating risks associated with LLM misuse.
• Balancing LLM safety and utility is achievable with advanced defense mechanisms.