# [Compressive Privacy Generative Adversarial Network](https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8963921)
The authors propose a generative model for generationg representations that retain utility and can defend against reconstruction attacks.

### Proposed Method :

- The model privatizes the data and feeds it to the reconstructor and service.
- The loss is modeled such that there is balance between utility and privacy.
- The funnel layers are use in privatisation mechanism to ensure the low dimensionailty of the outputs.
- Different reconstruction schemes are evaluated neural nets,linear ridge regression and kernel ridge regression.

### Experiments :
- The model is evaluated on both synthetic and real datasets.
- These are MNIST,HAR,GENKI-4K,CIFAR,SVHN and CelebA.
- The model is compared with Noise,DNN,DNN(resize) and RAN.
- The results show that the proposed model provides a good privacy/utility tradeoff while maintaining accuracy.
- The privacy/utility is controlled using the funnel layer
**Improvements:**
- The improvements to be noted are:
- Light-weighted Privatiser
- Better quantitative metric
- Privacy problem: As the method uses raw data for training it is a cause for concern.
- Optimisation of GAN objective function