An Adversarial Approach to Private Flocking in Mobile Robot Teams

Abstract

Privacy is an important facet of defence against adversaries. In this letter, we introduce the problem of private flocking. We consider a team of mobile robots flocking in the presence of an adversary, who is able to observe all robots' trajectories, and who is interested in identifying the leader. We present a method that generates private flocking controllers that hide the identity of the leader robot. Our approach towards privacy leverages a data-driven adversarial co-optimization scheme. We design a mechanism that optimizes flocking control parameters, such that leader inference is hindered. As the flocking performance improves, we successively train an adversarial discriminator that tries to infer the identity of the leader robot. To evaluate the performance of our co-optimization scheme, we investigate different classes of reference trajectories. Although it is reasonable to assume that there is an inherent trade-off between flocking performance and privacy, our results demonstrate that we are able to achieve high flocking performance and simultaneously reduce the risk of revealing the leader.

Human vs. Deep Neural Network Performance at a Leader Identification Task

“…As this suggests, the leader identification decision is difficult and for a human would depend on complex judgments based on factors such as the Gestalt principle of Common Fate (Sturzel & Spillman, 2004). Zheng et al (2020) published the first evaluation of a leader hiding algorithm. They started with a swarm following the simple Reynolds (1987) flocking algorithm described above in which a leader led the swarm through a sequence of waypoints.…”

“…They started with a swarm following the simple Reynolds (1987) flocking algorithm described above in which a leader led the swarm through a sequence of waypoints. Zheng et al (2020) developed what they called privacy preserving flocking using an alternating optimization procedure with genetic algorithms in which controller parameters were optimized for preserving privacy during flocking optimization, while leader identification accuracy was optimized in the following discrimination learning phase. While the discriminator, termed the adversary, grew increasingly more powerful over the course of their experiment the privacy preserving flocking algorithm also improved and was particularly effective in preserving privacy from the discriminator over curvilinear paths.…”

“…In this paper we describe the development of private flocking using multi-agent reinforcement learning (MARL) in an adversarial framework similar to that of Zheng et al (2020). Flocking swarm members and their adversary were trained in phases: first the swarm to follow the leader to the goal, then the adversary to identify that leader, and finally the swarm to hide its leader from a learning adversary.…”

Fixed-time leader-following flocking and collision avoidance of multi-agent systems with unknown dynamics

“…In recent years, the flocking control of multi-agent systems (MAS) has attracted a considerable amount of attention. Many efforts have been paid to understand how a school of fish, a crowd of people, a group of birds, and a swarm of bacteria can cluster without centralized coordination (Liu and Jiang, 2020;Mu and He, 2019;Zheng et al, 2020). In the real-world problems, learning the mechanism of flocking in biological groups is very important to the development of many artificial autonomous systems such as crowd evacuation, transport management, and formation control of the intelligent vehicles, the unmanned air vehicles or the autonomous surface vehicles (Dong et al, 2014;Hu et al, 2018).…”

Exposing Spoofing Attack on Flocking-Based Unmanned Aerial Vehicle Cluster: A Threat to Swarm Intelligence

“…Privacy leak and collision are two common security issues for the UAV cluster. In the works[32][33][34][35], researchers studied privacy issues in UAV clusters, such as privacy refers to preventing the inference of the leader's identity in leader-follower structure swarms. Our work belongs to the latter security issue of flocking-based cluster and collision.…”