Hyperbolic Safety-Aware Vision-Language Models

2025-03-15

AI Summary

Vision-language models (e.g., CLIP) struggle to safely discriminate unsafe content during retrieval, posing critical safety risks. Method: Departing from conventional “knowledge forgetting” paradigms, this work proposes the first hyperbolic partially ordered representation framework for safety-aware multimodal learning. It constructs an asymmetric entailment encoding scheme between safe and unsafe content to enable hierarchical discrimination and dynamic response; integrates hyperbolic embedding, entailment loss, and CLIP architecture extension to jointly optimize multimodal safety classification and retrieval. Contribution/Results: The method preserves original retrieval performance while significantly improving unsafe content detection accuracy. It supports runtime interpretable and controllable safety-policy switching. Evaluated on multiple benchmarks, it achieves state-of-the-art trade-offs between safety and utility.

Technology Category

• Computer Vision: Multi-modal Vision
• Natural Language Processing: Safety and Robustness
• Machine Learning: Multimodal Learning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Addressing the retrieval of unsafe content from vision-language models such as CLIP is an important step towards real-world integration. Current efforts have relied on unlearning techniques that try to erase the model's knowledge of unsafe concepts. While effective in reducing unwanted outputs, unlearning limits the model's capacity to discern between safe and unsafe content. In this work, we introduce a novel approach that shifts from unlearning to an awareness paradigm by leveraging the inherent hierarchical properties of the hyperbolic space. We propose to encode safe and unsafe content as an entailment hierarchy, where both are placed in different regions of hyperbolic space. Our HySAC, Hyperbolic Safety-Aware CLIP, employs entailment loss functions to model the hierarchical and asymmetrical relations between safe and unsafe image-text pairs. This modelling, ineffective in standard vision-language models due to their reliance on Euclidean embeddings, endows the model with awareness of unsafe content, enabling it to serve as both a multimodal unsafe classifier and a flexible content retriever, with the option to dynamically redirect unsafe queries toward safer alternatives or retain the original output. Extensive experiments show that our approach not only enhances safety recognition but also establishes a more adaptable and interpretable framework for content moderation in vision-language models. Our source code is available at https://github.com/aimagelab/HySAC.

Problem

• Retrieval of unsafe content in vision-language models
• Limitations of unlearning techniques in content discernment
• Enhancing safety and adaptability in content moderation

Innovation

• Hyperbolic space encodes safe and unsafe content.
• Entailment loss models hierarchical relations in CLIP.
• Dynamic redirection of unsafe queries to safer alternatives.