This paper proposes a novel membership inference attack (MIA) method to determine whether a specific image is included in the training data of a diffusion model, addressing copyright issues in AI-generated images. To overcome the limitation of existing MIA methods requiring access to the internal U-net of the model, we propose a method that uses only the image-to-image transformation API to determine whether the training data is included without accessing the internal structure of the model. This method leverages the fact that the model can more easily obtain noisy predictions for the training data. Therefore, the method averages the results using the API multiple times and compares them with the original image. Experiments on DDIM, Stable Diffusion, and Diffusion Transformer architectures demonstrate that our method outperforms existing methods.
