Pando: Efficient Byzantine-Tolerant Distributed Sensor Fusion using Forest Ensembles

Ad hoc communication networks provide a robust and low-power method for sensors to return their observations to an authority. However, ensuring that the returned values accurately represent the environment being sensed poses challenges. In particular, malicious sensors controlled by an adversary may collude to return systematically misleading or falsified results. Previous work examines this challenge for weak adversaries and under strong assumptions, limiting practical applicability. In this work, we propose Pando, a novel approach for mitigating the Byzantine distributed sensor fusion problem. We introduce an approach for the decentralized creation of a forest ensemble, made up of overlapping, hierarchical message passing routes that provide robust and efficient delivery while also mitigating adversarial interference. We then leverage a modified, homomorphic Merkle tree structure to create a novel Byzantine-tolerant message passing protocol. Finally, we propose a classification-informed data fusion algorithm based on these methods. We evaluate the correctness and efficiency of our approach under several attack models, and discuss functionality improvements over the state of the art.