Joonsung presented his paper, “No Caption, No Problem: Caption-Free Membership Inference via Model-Fitted Embeddings,” at the ICLR 2026 Conference, held from April 23rd – 27th, 2026, in Rio de Janeiro, Brazil.
He introduced MoFit, the first caption-free membership inference attack for diffusion models, which determines whether an image was used during training without access to its original caption. The key idea is that training samples are more sensitive to mismatched conditioning than non-members, so the method optimizes a model-fitted surrogate image and embedding to amplify this difference during denoising. Experiments on multiple Stable Diffusion models show that MoFit consistently outperforms prior VLM-caption-based attacks and can even surpass some methods that use ground-truth captions.