Do Not Perturb Me: A Secure Byzantine-Robust Mechanism for Machine Learning in IoT

Abstract

With the development of Internet of Things (IoTs) and big data, collaborative machine learning has achieved many impressive successes in IoT to improve the system performance and provide more diversified services for people. Although one of the motivations for collaborative learning is privacy preservation, the adversary can still launch an inference attack through task nodes' shared information. What's worse, a few task nodes may perform as a Byzantine attacker to compromise the entire system. Many Byzantine-robust mechanisms have been proposed, but they relied on outsourcing the calculation on two non-colluding servers which were not realistic in practice or had privacy issues in one-server architecture. In this paper, we design a novel mechanism for secure Byzantine-robust collaborative machine learning, namely Omega, to allow IoT devices to achieve the collaborative model training without exposing their local data to the others. Specifically, we construct a single-server architecture to achieve the private aggregation of parameter gradients, which protects task nodes' local data even n-1 of n nodes colluded. A new secure Byzantine-robust protocol is also designed to resist the Byzantine attack and this protocol can be extended to any distance-based robust rule. Furthermore, we prove that Omega can ensure task nodes' privacy preservation. Finally, we conduct an experiment to evaluate Omega over real-world dataset and empirical results demonstrate that Omega can efficiently achieve the collaborative machine learning.