UniNDM: A Unified Noise-driven Detection and Mitigation Framework Against Sexual Content in Text-to-Image Generation

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Text-to-image (T2I) diffusion models exhibit vulnerabilities to implicit sexual prompts, generating inappropriate content due to model biases.
• Current safety mechanisms struggle to detect subtle malicious intent and adapt to diverse implicit scenarios.

Purpose of the Study:

• To propose UniNDM, a unified noise-driven framework for robust safety in T2I diffusion models.
• To address limitations in detecting and mitigating implicit sexual content in AI-generated imagery.

Main Methods:

• Leveraging the early-stage noise separability between normal and explicit content in diffusion processes.
• Developing a lightweight noise-based detector and noise-enhanced adaptive negative guidance for mitigation.
• Extending the framework to U-Net and Diffusion Transformer (DiT) architectures.

Main Results:

• Achieved superior accuracy in detecting implicit sexual content with minimal computational overhead.
• Demonstrated substantial improvements over state-of-the-art methods on natural and adversarial datasets.
• Successfully extended the framework to various diffusion model architectures, including SDv1.4-XL and SDv3.

Conclusions:

• UniNDM offers a unified and effective approach to enhance safety in T2I diffusion models against implicit harmful content.
• The noise-driven framework provides a promising direction for future AI safety research in generative models.