uds-lsv / sidp

Robust Differentially Private Training of Deep Neural Networks

SIDP

The PyTorch implementation of 'Robust Differentially Private Training of Deep Neural Networks'. https://arxiv.org/abs/2006.10919

Usage

To run the code on MNIST and CIFAR-10 datasets, execute vision.py with desired parameters:

python vision.py 

For the text classification model based on the AGNews Corpus:

python main_text_class.py [experiment_name] [noise_std] [clip]
python main_text_class.py sidp 0.3 7

Results

MNIST

Privacy epsilon	∞	7	3	1	0.5	0.1	0.05	0.025
DPSGD (LeNet5)	99.2	97	96.34	94.11	91.1	83.0	78.96	31.56
SI-DPSGD (LeNet5)	99.2	98.9	98.9	98.72	99.1	99.0	98.84	90.82
SI-DPSGD (BN-LeNet5)	99.2	99.17	99.17	99.15	99.18	99.14	99.12	98.58

CIFAR-10

Privacy epsilon	∞	8	4	2	1	0.5	0.1	0.05
DPSGD (TF-tutorial)	80.0	73.0	70.0	67.0	NA	NA	NA	NA
SI-DPSGD (TF-tutorial)	80.0	78.10	77.70	76.0	76.05	74.20	73.80	74.05
SI-DPSGD (ResNet-18)	93.50	90.20	90.16	90.26	90.09	89.67	84.88	84.47

AGNews Text Classification

Privacy epsilon	∞	7	3	1	0.5	0.1	0.05
DPSGD (BiLSTM-DL)	88.5	83.9	80.0	81.1	77.9	37.5	31.8
SI-DPSGD (BiLSTM-DL)	88.5	85.9	85.7	83.3	81.2	77.9	56.7
DPSGD (LN-BiLSTM-DL)	88.5	83.5	82.4	82.0	78.9	50.1	31.6
SI-DPSGD (LN-BiLSTM-DL)	88.5	87.8	87.6	85.7	85.4	84.3	80.1

References

• PyTorch implementation of DPSGD is taken from https://github.com/ebagdasa/pytorch-privacy.