Amplifying Membership Signal Through Chained Regeneration

Imagine you're building a system that generates content, like text or audio, based on a large dataset. One major concern is that these systems can memorize specific pieces of training data, which raises privacy issues and copyright concerns. If someone can prove that a model is generating content that includes their copyrighted material, it could lead to serious legal problems. This is where sample verification becomes crucial — you want to know if a specific output was generated based on a training example or if it’s entirely new. However, current methods for checking this, known as membership inference attacks, often rely on generating outputs just once. This one-shot approach can be weak and doesn’t provide enough information to make solid conclusions, especially when dealing with different types of data like text, images, or audio. This limitation is what's called weak signal detection in membership inference attacks. The authors of this paper propose a new framework called MADreMIA, which takes a different approach. Instead of generating outputs just once, it uses a process where each output becomes the input for the next generation. This iterative method allows the system to gather more information and improve the accuracy of its membership evidence. The key insight is that memorized samples show more coherence and degrade more slowly than non-member samples during this iterative process. Practically, this means that if you're working with generative models, using MADreMIA could help you better identify whether your outputs are infringing on someone else's data, making your system more robust against privacy violations.

The framework introduces a new approach to enhance membership inference attacks across multiple modalities.

The claims are supported by comprehensive evaluations across various model families and modalities.