AI Secure LLM Safety Leaderboard

Hugging Face and the Secure Learning Lab have introduced the LLM Safety Leaderboard, a specialized evaluation platform designed to assess the trustworthiness of Large Language Models (LLMs). Powered by the DecodingTrust framework, the leaderboard provides a standardized way to measure safety risks across both open and closed models to ensure they meet regulatory and safety requirements before real-world deployment.

The DecodingTrust Framework

DecodingTrust is a comprehensive evaluation platform that assesses LLM trustworthiness through eight distinct perspectives. The framework is designed to provide a holistic view of model safety, utilizing novel red-teaming algorithms to stress-test models in diverse scenarios.

Trustworthiness Dimensions

  • Toxicity: Evaluated using optimization algorithms and prompt generative models to create challenging user prompts, alongside 33 specific system prompts (including role-play and task reformulation) to test toxicity scores via the Perspective API.
  • Stereotype Bias: Measured by testing 24 demographic groups and 16 stereotype topics across three prompt variations, with scores averaged over five prompts per topic.
  • Adversarial Robustness: Assessed by applying five adversarial attack algorithms against open models (Alpaca, Vicuna, and StableVicuna) and testing other models using the resulting adversarial data across five tasks.
  • OOD (Out-of-Distribution) Robustness: Evaluated by transforming input styles (e.g., into Shakespearean or poetic forms) or testing knowledge that is absent from the model's training data.
  • Robustness against Adversarial Demonstrations: Tested using demonstrations containing misleading information, such as backdoor attacks, spurious correlations, and counterfactual examples.
  • Privacy: Assessed at three levels: privacy leakage from pretraining data, privacy leakage during conversations, and the model's understanding of privacy-related words and events.
  • Ethics: Evaluated using the ETHICS and Jiminy Cricket datasets to design jailbreaking systems and user prompts that test the model's ability to recognize immoral behavior.
  • Fairness: Measured by controlling protected attributes across different tasks to generate challenging questions in both zero-shot and few-shot settings.

Key Research Findings

Research conducted through the DecodingTrust framework revealed several critical vulnerabilities in current LLMs:

  • Model Vulnerability: GPT-4 was found to be more vulnerable than GPT-3.5 in certain contexts.
  • Lack of Consistent Performance: No single LLM consistently outperforms all others across every trustworthiness perspective.
  • Performance Trade-offs: There are inherent trade-offs between different trustworthiness perspectives.
  • Sensitivity to Prompting: LLMs are vulnerable to adversarial or misleading prompts. Specifically, privacy leakage can vary based on phrasing; for example, GPT-4 may not leak information when prompted with "in confidence" but may do so when prompted with "confidentially."

Model Submission Process

Developers can submit their models to the LLM Safety Leaderboard via the "Submit here!" panel on the leaderboard space. To be eligible for evaluation, models must meet the following criteria:

  1. Format: Model weights must be converted to the safetensors format for improved safety and loading speed.
  2. Accessibility: Models must be public.
  3. Compatibility: Models must be loadable using Hugging Face AutoClasses (AutoConfig, AutoModel, AutoTokenizer).
  4. Current Limitations: Models requiring use_remote_code=True are not currently supported.

Sources