Threat-Resilient Decentralized Online Learning

Decentralized online learning (DOL) is a computing paradigm where multiple agents cooperate to predict online data samples via a decentralized ommunication network. Despite its superior capability in handling large-scale decentralized online data, it has yet to be widely adopted in real-world applications due to its proneness to endogenous and exogenous threats. Endogenous threat is caused by the adversarial agents in the communication network, which share arbitrary models with their neighbors. Exogenous threat is originated from entities other than the agents themselves, including malicious data generators and erroneous feedback providers that present manipulated data features or potentially erroneous feedbacks to the agents during the DOL process. Our preliminary evaluation results illustrate that the performance of the state-of-the-art DOL strategies is significantly weakened by such threats (see
Fig. 1). These threats have yet to be adequately addressed in the literature on DOL as well as adversarial machine learning techniques.
This research will thoroughly investigate the problem of endogenous and exogenous threats in DOL and attempt to derive suitable solutions (with performance guarantee) to the problem. A central challenge of the problem is that negative impacts caused by threats on model learning or online predictions can propagate through the communication network from agent to agent, which complicates the design of a threat-resilient DOL approach.
Our objectives are threefold: (1) To investigate countermeasures against endogenous and exogenous threats, respectively. (2) Based on understanding of the strategies against each type of threat and their interactions, to design a threat-resilient decentralized online learning solution with performance guarantee. (3) To build a simulator and a testbed for performance evaluation of our proposed algorithms using real-world datasets and workloads.
This project is the first one that considers both endogenous and exogenous threats in decentralized online learning and will develop a new algorithmic framework for threat-resilient DOL with performance guarantee. Besides, we will validate the effectiveness and efficiency of the proposed solutions by application to decentralized network traffic classification at the network edge. Our research will provide new theoretical principles in the field of computer science and related disciplines. It will promote the successful deployment of reliable DOL in critical real-world applications, which can benefit high-tech companies of different scales in Hong Kong, the Greater Bay Area, and other areas.