arxivcs.CV2026-07-05
TrustCLIP: Learning Private Visual Features via Adversarial Reconstruction
Nikos Athanasiou, Ilya A. Petrov, Angela Yao, Shugao Ma, Eric Sauser, Edoardo Remelli, et al.
Vision and vision-language models rely on high-level visual representations that are increasingly used across recognition, retrieval, and multimodal reasoning pipelines. However, recent advances in generative modeling have shown that such features can often be inverted, enabling…