A PRIVACY PRESERVING VISUAL RECOGNITION THROUGH CONTRADICTORIC LEARNING

摘要

A method for protecting visual private data by preventing data recovery from latent representations of deep networks is presented. The method includes obtaining latent features (316) from an input image (312) and learning, via an adversarial recovery learning framework (318), from privacy-preserving feature representations to maintain utility performance and prevent data recovery by: simulating a black box model inversion attack by training a decoder (332) to recover the input image from the latent features; and training an encoder (314) to maximize recovery error to prevent the decoder from recovering the latent features inverted while minimizing the loss of tasks.