Opacus-DPCR is a library extend the widely used Opacus for learning with differential privacy (DP). Compared with Opacus, our Opacus-DPCR introduces differential privacy continuous data release (DPCR) to improve learning accuracy. It integrates various DPCR models (including our proposed BCRG and ABCRG). For more details, please refer to our work "A Federated Learning Framework Based on Differential Privacy Continuous Data Release"[1].

You can install Opacus-DPCR by

• pip install dpcrpy
• pip install opacus_dpcr

In the demo folder, we provide two sample codes for testing DPCR (dpcrTest.py) and privacy learning with DPCR (opacusDpcrTest.py).

In Opacus-DPCR, we support multiple DPCR models for building parametric models. They are as follows:

1. SimpleMech: Essentially the same as Opacus. They both add noise after each gradient calculation and then accumulate the noisy gradients to construct the parametric models.
2. TwoLevel: The method proposed in [2]. It is a primary DPCR model.
3. BinMech: The method proposed in [2]. It is a classical DPCR model, applying a binary tree to achieve a logarithmically growing RMSE.
4. FDA: The method proposed in [3]. It first applies binary indexed tree (BIT) on DPCR for higher accuracy. But it is only suitable for Laplacian noise.
5. BCRG: The method proposed in [1]. It designs the optimization method for BIT-based DPCR with Gaussian noise.
6. ABCRG: The method proposed in [1]. It improves the BCRG by using the residual sensitivity to further boost the accuracy.

Our Opacus-DPCR keeps a high compatibility with Opacus. You only need to easily modify a few lines of code for applying our achievements.

As an instance, for the classic Opacus example from (https://github.com/pytorch/opacus/blob/main/examples/mnist.py), you just do the following modifications to apply Opacus-DPCR as shown in the following Figure. For more details, see our Example Code.

[1] J. Cai, X. Liu, Q. Ye, Y. Liu and Y. Wang, “A Federated Learning Framework Based on Differential Privacy Continuous Data Release.”

[2] T.-H. H. Chan, E. Shi, and D. Song, “Private and continual release of statistics,” ACM Transactions on Information and System Security (TISSEC), vol. 14, no. 3, pp. 1–24, 2011.

[3] J. Cai, W. U. Yingjie, and X. Wang, “Method based on matrix mechanism for differential privacy continual data release,” Journal of Frontiers of Computer Science and Technology, 2016.