Skip to content
Switch branches/tags
Go to file

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


Caffe implementation of Deep Private-Feature Extraction (DPFE) (arXiv).

What is a Private-Feature? extracted feature of the input data which is non-informative with respect to the user sensitive information, while still useful for other desired tasks. Private-feature is useful in the process of user-cloud interaction, when the user provides his/her data to a cloud service and also has some privacy concerns.

Here we consider the problem of facial attributes (like gender and age) prediction, in which we want to use a cloud service to tag an image with facial attributes and at the same time we don't want to reveal our identity to the service provider.

Considering a pretrained deep network for facial attributes prediction (original model) first we should choose an intermediate layer of this network and then we should reduce its dimensionality. This can be done by using the file code/ Our theory is based on cross entropy loss and batch normalization regularizer so we should fine the models: code/ Now its time to create simple or DPFE models. Considering an attribute set (like {gender & age}) we should fine tune each model for the chosen set with or without contrastive loss (for simple and dpfe). This is done by first using code/ and then using code/ and code/ Now we have all models ready and we should evaulate them b creating the accuracy-privacy curve. We introdcued the log-rank privacy measure in the paper and by using the file code/ we can evaluate all models. At the end code/ help us for creating the plots.

Be aware that we used CAFFE and you should use the file code/ as a new python layer. All results will be saved in the result directory. Also read the comments at the first of each python file before running it.


Implementation of "Deep Private-Feature Extraction (DPFE)"




No releases published


No packages published