A group of researchers are warning that AI image models, such as DALL-E 2, are vulnerable to being “tricked” into regenerating their training images, which poses a significant privacy concern, especially in applications such as medical imaging.
The team, which includes researchers from various institutions such as Google Brain, Princeton University, and University of California Berkeley, discovered that generative diffusion models memorize and recreate their training data. This raises questions about data privacy, digital forgery, and how these models should be responsibly deployed.
The researchers highlight the need to understand the risks of generative models before applying them to private domains, such as medical research, and advise researchers and practitioners to be cautious when training on uncurated public data without first addressing ethical and privacy implications.
