Byzantine-Tolerant Inference in Distributed Deep Intelligent System: Challenges and Opportunities

Cyber-physical systems in the Internet-of-Things (IoT) era increasingly need responses that are not only timely, but also intelligent. To that end, decentralized Deep Neural Network (DNN) systems have been studied for near-sensor processing to enable localized inference and global network partition in a given, limited power budget. In the real world, however, such systems are fundamentally heterogeneous in per-node information quality, since each virtually or physically dispersed sensor captures only part of the observable environment. Recent work in collective DNN systems has leveraged this variation in information quality across nodes in a network hierarchy to improve accuracy and reduce design cost. Unfortunately, in noisy or dynamic environments, there are clear optimization challenges in properly measuring information quality or its proxies, which can make such systems even more sensitive to noise than a quality-agnostic design. This implies that the unequal weighting in an information-quality exploiting distributed DNN will likely present new opportunities to subvert collective decision making. Inspired by the classical fault models of distributed systems, specifically those designed to endure coordinated node failure (Byzantine failure), this article explores interactions between collaborative DNN models, their reliance on information quality metrics in various fault scenarios, and the corresponding system design costs for achieving reasonable accuracy and reliability. As a proof-of-concept, we perform a case study on a distributed multi-view camera system operating under faults introduced both by environmental noise and adversarial inputs, and present results on inference robustness supported by consensus mechanisms in Byzantine settings.