Towards Privacy Aware Deep Learning for Embedded Systems (ACM SAC'22, PPML@NeurIPS'20)

This is the code repository for the paper "Towards Privacy Aware Deep Learning for Embedded Systems".

Experiments

All the code is in Jupyter notebooks for easy reproducibility. Following are the folders and their contents:

• Quantization: Contains the privacy risk analysis for binarization and XNOR networks.
• StdArchitectures: Contains the privacy risk analysis for standard deep learning architectures designed for efficiency such as SqueezeNet and MobileNet.
• Pruning: Contains the privacy risk analysis of pruning the models followed by retraining. "Sparsity" folder provides an alternate implementation in a different ML library.
• Defences: Contains the code for blackbox defences such as adversarial regularization and differential privacy for comparison with Gecko models.
• Knowledge Distillation: Contains the code for homogeneous and heterogenous knowledge distillation of quantized models which forms the Gecko training methodology.

Credits

The code for binarization is adapted from https://github.com/itayhubara/BinaryNet.pytorch and https://github.com/jiecaoyu/XNOR-Net-PyTorch. Pruning code adapted from https://github.com/yangbingjie/DeepCompression-PyTorch. MIA attack code adapted from https://github.com/inspire-group/privacy-vs-robustness.