Recoverable Privacy-Preserving Image Classification through Noise-like Adversarial Examples

This is the official code for the paper titled as "Recoverable Privacy-Preserving Image Classification through Noise-like Adversarial Examples", accepted by ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM).

Please download the datasets and the weight of RIC pre-trained on VGGFace2 or SVHN datasets from GoogleDrive.

Put the model weights to the folder "Weights"

For SVHN: The download process of the dataset will be automated.

For VGGFace2: Plead download the "VGGFace2_vggface2_train.tar" subset from this url, then name the extracted folder "train" and place it in the "data/vggface2" path.

Take the training on the SVHN dataset for an example:

1. Go to SVHN.py, set the variable stage='train'.
2. Run SVHN.py

Take the test on the SVHN dataset for an example,

1. Go to SVHN.py, set the variable stage='test'.
2. Run SVHN.py

Run ImageNet.py or Cifar10.py directly.