Towards Realistic Guarantees: A Probabilistic Certificate for SmoothLLM

2025-11-23

AI Summary

SmoothLLM provides certified defenses against jailbreak attacks but relies on the strong and practically unrealistic “k-instability” assumption, undermining the credibility of its safety guarantees. Method: We propose a more realistic probabilistic safety definition—(k, ε)-instability—requiring that model outputs remain stable with probability at least 1−ε under up to k perturbations. Building upon this, we develop a data-driven framework to estimate certified safety lower bounds, integrating empirical attack success rate modeling with SmoothLLM’s smoothing mechanism. The framework supports robustness certification against mainstream gradient- and semantics-based jailbreak attacks, including GCG and PAIR. Contribution/Results: Our approach ensures theoretical rigor while maintaining deployment feasibility. Evaluated across multiple LLMs and attack scenarios, it significantly improves both certification coverage and robustness, delivering verifiable, practical security guarantees for large language models.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

The SmoothLLM defense provides a certification guarantee against jailbreaking attacks, but it relies on a strict `k-unstable' assumption that rarely holds in practice. This strong assumption can limit the trustworthiness of the provided safety certificate. In this work, we address this limitation by introducing a more realistic probabilistic framework, `(k, $varepsilon$)-unstable,' to certify defenses against diverse jailbreaking attacks, from gradient-based (GCG) to semantic (PAIR). We derive a new, data-informed lower bound on SmoothLLM's defense probability by incorporating empirical models of attack success, providing a more trustworthy and practical safety certificate. By introducing the notion of (k, $varepsilon$)-unstable, our framework provides practitioners with actionable safety guarantees, enabling them to set certification thresholds that better reflect the real-world behavior of LLMs. Ultimately, this work contributes a practical and theoretically-grounded mechanism to make LLMs more resistant to the exploitation of their safety alignments, a critical challenge in secure AI deployment.

Problem

• SmoothLLM defense relies on unrealistic k-unstable assumption
• Strong assumption limits trustworthiness of safety certificates
• Need probabilistic framework for realistic jailbreak attack certification

Innovation

• Probabilistic framework certifies defenses against jailbreak attacks
• Data-informed lower bound improves trustworthiness of safety certificates
• Realistic certification thresholds reflect real-world LLM behavior