eduzhai > Applied Sciences > Engineering >

Image Transformation Network for Privacy-Preserving Deep Neural Networks and Its Security Evaluation

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 4 pages

Abstract: We propose a transformation network for generating visually-protected imagesfor privacy-preserving DNNs. The proposed transformation network is trained byusing a plain image dataset so that plain images are transformed into visuallyprotected ones. Conventional perceptual encryption methods have a weakvisual-protection performance and some accuracy degradation in imageclassification. In contrast, the proposed network enables us not only tostrongly protect visual information but also to maintain the imageclassification accuracy that using plain images achieves. In an imageclassification experiment, the proposed network is demonstrated to stronglyprotect visual information on plain images without any performance degradationunder the use of CIFAR datasets. In addition, it is shown that the visuallyprotected images are robust against a DNN-based attack, called inversetransformation network attack (ITN-Attack) in an experiment.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×