1: \begin{abstract}

2:     We study differentially private (DP) machine learning

3:     algorithms as instances of noisy fixed-point iterations, in order to

4:     derive privacy and utility results from this well-studied framework. We

5:     show that this new perspective recovers popular private gradient-based

6:     methods like DP-SGD and provides a principled way to design and

7:     analyze new

8:     private optimization algorithms in a flexible manner. Focusing on the

9:     widely-used Alternating Directions Method of Multipliers (ADMM) method, we

10:     use our general framework to derive novel private ADMM algorithms for

11:     centralized, federated and fully decentralized learning. For these three algorithms, we establish

12:     strong privacy guarantees leveraging privacy

13:     amplification by iteration and by subsampling.

14:     % We prove

15:     % strong DP guarantees by relying on 

16:     % We also introduce a

17:     % federated ADMM algorithm compatible with

18:     % client subsampling, which simultaneously provides local DP guarantees and

19:     % stronger bounds for central DP.

20:     Finally, we provide utility guarantees using a unified analysis that exploits a recent

21:     linear convergence result for noisy fixed-point iterations.

22: \end{abstract}

23: