RI: Small: Removing and Preventing Harmful Generations in AI Image Generators

Purdue University, West Lafayette IN

Investigators

Abstract

Open-source releases of AI Image Generators have dramatically lowered the entry barrier to generating image content. Using commercial-grade hardware, anyone can generate realistic and aesthetically pleasing images with a few lines of code, or even no code. With this technological advancement, there is also an increasing responsibility to ensure that the technology does more good than harm to society. For example, AI image generators could be used to copy artists' work without their permission or create fake celebrity images. This project aims to develop techniques that prevent these harmful use cases, promoting a safer digital environment and mitigating AI image generator misuse. Additionally, the project will educate the public about the risks associated with AI image generators and raise awareness about privacy and the potential dangers of sharing their photograph data. To achieve these goals, the project categorizes harmful use cases into unintentionally and intentionally developed harmful capabilities. To address unintentionally harmful capabilities, the project develops methods to remove content from a trained AI generator (Thrust 1). Additionally, to prevent intentionally developed harmful capabilities, the project will develop techniques to "immunize" the AI generators (Thrust 2). This means it would be more challenging for a trained model to be fine-tuned on harmful concepts, e.g., formulated as a bi-level optimization framework. Finally, the project aims to efficiently solve the formulated optimization problems in Thrust 1 and Thrust 2. This includes studying the techniques and models that are well-suited to distributed modes of computation, as well as leveraging underlying structural properties, such as sparsity, to achieve further performance gains. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.