RIB-Guard: A Risk-Aware Information Bottleneck Defense for Black-Box Large Language Models

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Cybersecurity

Background:

• Large language models (LLMs) are susceptible to jailbreak attacks, particularly in black-box scenarios.
• Existing defenses often require white-box access or rely on the model's inherent alignment, limiting their applicability.
• Information bottleneck methods frame prompt protection as a compression task but face optimization challenges.

Purpose of the Study:

• To develop a novel defense mechanism, RIB-Guard, for enhancing the security of black-box large language models against jailbreak attacks.
• To address the limitations of existing information bottleneck defenses by enabling black-box optimization and model-agnostic safety guidance.
• To create a framework that balances prompt compactness, utility preservation, and residual risk reduction.

Main Methods:

• Proposed RIB-Guard, a safety-aware information bottleneck defense specifically designed for black-box LLMs.
• Implemented a token-level masking policy learned via reinforcement learning using only black-box feedback.
• Introduced an independent lightweight safety head to estimate residual jailbreak risk and provide model-agnostic safety guidance.

Main Results:

• Demonstrated improved jailbreak robustness in direct single-turn harmful and benign prompt settings.
• Showcased competitive preservation of benign utility alongside enhanced security.
• Successfully extended information bottleneck-based prompt protection from white-box to black-box environments.

Conclusions:

• RIB-Guard offers a significant advancement in defending black-box LLMs against jailbreak attacks.
• The proposed framework provides a step towards safety-aware, information-theoretic front-end defenses for LLMs.
• This research contributes to more secure and reliable deployment of large language models.