1: \begin{abstract}

2: Decentralized optimization enables a network of agents to

3: cooperatively optimize an overall objective function without a

4: central coordinator and is gaining increased attention in domains as

5: diverse as control, sensor networks, data mining, and robotics.

6: However, the information sharing among agents in decentralized

7: optimization also  discloses  agents' information, which is

8: undesirable or even unacceptable when involved data are  sensitive.

9: This paper proposes two gradient based decentralized optimization

10: algorithms that can protect participating agents' privacy without

11: compromising optimization accuracy or incurring heavy

12: communication/computational overhead. This is in distinct difference

13: from differential privacy based approaches which have to trade

14: optimization accuracy  for privacy, or encryption based approaches

15: which incur heavy communication and computational overhead. Both

16: algorithms leverage a judiciously designed mixing matrix and

17: time-varying uncoordinated stepsizes to enable privacy, one using

18: diminishing stepsizes while the other   using non-diminishing

19: stepsizes. Both algorithms only require a participating agent to

20: share {\it one} message with a neighboring agent in each iteration

21: to reach convergence  to an exact optimal solution, which is in

22: contrast to most gradient-tracking based algorithms requiring every

23: agent to share two messages (an optimization variable and an

24: auxiliary gradient-tracking variable)

25:   under non-diminishing stepsizes. Furthermore, both algorithms

26: can guarantee the privacy of a participating agent even when all

27: information  shared by the agent are accessible to an adversary, a

28: scenario in which

29:   most existing accuracy-maintaining privacy  approaches will fail to protect privacy. Simulation results confirm the effectiveness of the proposed

30: algorithms.

31: \end{abstract}

32: