English

TrustCLIP: Learning Private Visual Features via Adversarial Reconstruction

Computer Vision and Pattern Recognition 2026-07-05 v1

Abstract

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 realistic reconstructions of the underlying image and raising significant privacy risks. We revisit this problem through the lens of reconstruction and propose TrustCLIP, a reconstruction-driven framework that treats a feature-conditioned generator as an explicit privacy adversary. TrustCLIP learns a projection between encoder features and downstream modules that is explicitly optimized to degrade the reconstructions produced by generative attackers while retaining the necessary signals for downstream tasks. Unlike prior defenses that rely on discriminative privacy metrics, TrustCLIP directly optimizes against a generative reconstruction attacker, targeting a threat not captured by standard evaluation protocols. We demonstrate its effectiveness in both conventional classification and multimodal large language model pipelines. Across these settings, TrustCLIP consistently reduces the fidelity of generative inversions while maintaining downstream task performance. Project page: https://atnikos.github.io/trustclip/

Cite

@article{arxiv.2607.04484,
  title  = {TrustCLIP: Learning Private Visual Features via Adversarial Reconstruction},
  author = {Nikos Athanasiou and Ilya A. Petrov and Angela Yao and Shugao Ma and Eric Sauser and Edoardo Remelli and Shreyas Hampali and Johannes Schönberger and Fadime Sener and Bugra Tekin},
  journal= {arXiv preprint arXiv:2607.04484},
  year   = {2026}
}

Comments

https://atnikos.github.io/trustclip/