Securing Representations via Latent Disruption and Private Decoding

TMLR Paper6644 Authors

Abstract: Pre-trained encoders facilitate efficient data sharing through semantically rich latent embeddings, which, however, pose privacy risks under malicious inference or exploitation. We propose SEAL, an attack-agnostic framework that secures latent spaces by disrupting semantic dependencies based on information-theoretic principles. It prevents potential misuse while enabling selective reconstruction for trusted users. SEAL learns to encode controlled perturbations by minimizing the Matrix Norm-based Quadratic Mutual Information (MQMI) functional between original and secured embeddings within a hyperspherical latent space. Meanwhile, a private decoder, jointly trained with the SEAL encoder, ensures accurate reconstruction that is accessible only to authorized users. Extensive experiments on vision and text datasets demonstrate that SEAL effectively mitigates latent leakage, defends against inference attacks, and preserves reconstruction utility.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Feng_Liu2

Submission Number: 6644