Privacy-Preserving Distributed Kalman Filtering

Abstract

Distributed Kalman filtering techniques enable agents of a multiagent network to enhance their ability to track a system and learn from local cooperation with neighbors. Enabling this cooperation, however, requires agents to share information, which raises the question of privacy. This paper proposes a privacy-preserving distributed Kalman filter (PP-DKF) that protects local agent information by restricting and obfuscating the information exchanged. The derived PP-DKF embeds two state-of-the-art average consensus techniques that guarantee agent privacy. The resulting PP-DKF utilizes noise injectionbased and decomposition-based privacy-preserving techniques to implement a robust distributed Kalman filtering solution against perturbation. We characterize the performance and convergence of the proposed PP-DKF and demonstrate its robustness against the injected noise variance. We also assess the privacy-preserving properties of the proposed algorithm for two types of adversaries, namely, an external eavesdropper and an honest-but-curious (HBC) agent, by providing bounds on the privacy leakage for both adversaries. Finally, several simulation examples illustrate that the proposed PP-DKF achieves better performance and higher privacy levels than the distributed Kalman filtering solutions employing contemporary privacy-preserving techniques.

Event‐triggered differential private distributed fusion estimation for sensor networks with unknown cross‐correlations

“…Theorem 1. With utilizing the ET mechanism as Equation (6) to schedule transmissions of perturbed local estimates in the differential private distributed fusion estimation shown as Figure 2, the information for the fusion center obtained from sensor i at k instant is…”

“…Varieties formal notions of privacy and the corresponding privacy‐preserving fusion estimation methods have been proposed 4–6 . Among these, differential privacy 7,8 has gained more and more attentions.…”

Communication-Efficient and Privacy-Aware Distributed Learning

“…To further protect the node-sensitive information, at each consensus iteration m, agents share only a perturbed version of their public substate with their neighboring agents. Subsequently, at agent k, the decomposition and noise injection-based privacy-preserving ACF [27], at mth iteration is stated as…”

“…The HBC agent has access to it own local information {α k,(m) , β k,(m) , S n,(m) } as well as the shared information by its neighbors S n,(m) α l,(m) for l ∈ N − k . We can observe that agent privacy depends on the availability of the interaction and coupling weights for all agents at the HBC adversary [27]. Thus, to investigate the privacy leakage in the worst-case scenario, we also assume that the HBC agent has also access to the coupling and interaction weight matrices Θ and ∆.…”

“…It effectively cancels out the added noise during local aggregation steps, thereby enhancing the performance and privacy of the distributed learning algorithm. Meanwhile, a decomposition-based privacy technique makes inference more challenging for adversaries by randomly decomposing private information into two substates, of which only one is shared among agents [24]- [27]. In [28], a privacy-aware collaborative estimation strategy is proposed in the context of networked vehicles, wherein the vehicles transform their sensitive information into an alternative do-main prior to exchanging it with others.…”