Vipera: Towards systematic auditing of generative text-to-image models at scale

2025-03-14

AI Summary

Generative text-to-image (T2I) models pose systemic risks—including bias, offensiveness, and misinformation—yet existing auditing methods suffer from poor scalability, insufficient structural exploration, and inadequate coverage of auditing criteria. To address this, we propose the first scalable, systematic auditing framework for T2I models, integrating scene graph parsing, multimodal visual prompt design, LLM-driven intelligent recommendation of unexplored risk directions, and interactive visual analytics. Our framework supports hierarchical modeling and dynamic extension of auditing criteria, enabling structured, reproducible risk exploration. A user study with professional auditors demonstrates that our approach significantly improves organizational efficiency (+42%) and exploration breadth (+58%) across diverse risk criteria, while enhancing the systematicity, interpretability, and reproducibility of the auditing process.

Technology Category

• Natural Language Processing: Safety and Robustness
• Computer Vision: Diffusion Models for Vision
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Responsible Web: Audits of web technologies, standards, platforms, and applications

Abstract

Generative text-to-image (T2I) models are known for their risks related such as bias, offense, and misinformation. Current AI auditing methods face challenges in scalability and thoroughness, and it is even more challenging to enable auditors to explore the auditing space in a structural and effective way. Vipera employs multiple visual cues including a scene graph to facilitate image collection sensemaking and inspire auditors to explore and hierarchically organize the auditing criteria. Additionally, it leverages LLM-powered suggestions to facilitate exploration of unexplored auditing directions. An observational user study demonstrates Vipera's effectiveness in helping auditors organize their analyses while engaging with diverse criteria.

Problem

• Addresses risks in generative text-to-image models like bias and misinformation.
• Improves scalability and thoroughness of AI auditing methods.
• Facilitates structured exploration of auditing criteria using visual cues and LLM-powered suggestions.

Innovation

• Uses scene graphs for image collection sensemaking
• Leverages LLM-powered suggestions for exploration
• Facilitates hierarchical organization of auditing criteria