Controlled and Ethical Advancements in Diffusion-based Generative Models

The recent advancements in diffusion-based generative models have significantly pushed the boundaries of image editing and generation. A common theme across several papers is the focus on addressing unintended consequences and enhancing control over the generative process. Innovations in attribute leakage mitigation, concept erasure, and moderation of generalization in generative models are particularly prominent. These developments aim to balance the flexibility and power of diffusion models with the need for precise control and ethical considerations. Notably, methods for precise, fast, and low-cost concept erasure, as well as strategies to moderate the generalization of score-based models, are advancing the field by providing more robust and ethical generative solutions. Additionally, efforts to mitigate NSFW content generation and enhance model security through novel attack strategies and defense mechanisms are crucial for maintaining the integrity and safety of these models in real-world applications. These advancements collectively underscore a shift towards more controlled, secure, and ethically sound generative AI systems.

Sources

Addressing Attribute Leakages in Diffusion-based Image Editing without Training

Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters

Machine Unlearning Doesn't Do What You Think: Lessons for Generative AI Policy, Research, and Practice

Moderating the Generalization of Score-based Generative Model

Buster: Incorporating Backdoor Attacks into Text Encoder to Mitigate NSFW Content Generation

TraSCE: Trajectory Steering for Concept Erasure

Efficient Diversity-Preserving Diffusion Alignment via Gradient-Informed GFlowNets

Boosting Alignment for Post-Unlearning Text-to-Image Generative Models

Antelope: Potent and Concealed Jailbreak Attack Strategy

Built with on top of