1: \begin{abstract}

2: Decentralized stochastic optimization is the basic building block of

3: modern collaborative machine learning, distributed estimation and

4: control, and large-scale sensing. Since involved data  usually

5: contain sensitive information like user locations, healthcare

6: records and financial transactions,  privacy protection  has become

7: an increasingly pressing need in the implementation of decentralized

8: stochastic optimization algorithms. In this paper, we propose a

9: decentralized stochastic gradient descent algorithm which is

10: embedded with inherent privacy protection for every participating

11: agent against other participating agents and external eavesdroppers.

12: This proposed algorithm builds in a dynamics based

13: gradient-obfuscation mechanism to enable privacy protection without

14: compromising optimization accuracy, which is in significant

15: difference from differential-privacy based privacy solutions for

16: decentralized optimization that have to trade optimization accuracy

17: for privacy. The dynamics based privacy approach is encryption-free,

18: and hence avoids incurring heavy communication or computation

19: overhead, which is a common problem with encryption based privacy

20: solutions for decentralized stochastic optimization. Besides

21: rigorously characterizing the convergence performance of the

22: proposed decentralized stochastic gradient descent algorithm under

23: both convex objective functions and non-convex objective functions,

24: we also provide rigorous information-theoretic analysis of its

25: strength of  privacy protection. Simulation results for a

26: distributed estimation problem as well as numerical experiments for

27: decentralized   learning on a benchmark machine learning dataset

28: confirm the effectiveness of the proposed approach.

29: \end{abstract}

30: