METHOD FOR DETECTING ANOMALOUS LOCAL UPDATES AND ISOLATING MALICIOUS PARTICIPANTS IN FEDERATED LEARNING SYSTEMS

Abstract

This paper proposes a method for detecting anomalous local updates and isolating malicious participants in federated learning systems. The method is aimed at improving the resilience of distributed machine learning models to model poisoning attacks and distorted local updates. The proposed approach combines two protection levels: anomaly assessment of local updates based on their deviation from the collective update pattern and isolation of participants that repeatedly demonstrate suspicious behavior. An anomaly score is used to evaluate local updates, while an accumulated anomaly counter is applied to control long-term participant behavior. The paper presents the formalization of the proposed method, describes the adaptive anomaly threshold mechanism, and defines the procedure for forming a trusted participant set for further global model aggregation. An illustrative numerical example is provided to demonstrate the operation of the proposed approach and to show the possibility of detecting anomalous local updates while reducing their influence on the global model. A comparative analysis of the proposed approach with existing aggregation and malicious participant detection methods in federated learning systems is also presented. It is shown that, unlike conventional aggregation schemes, the proposed method provides not only suspicious update filtering but also control of repeated anomalous participant behavior. The proposed approach can be applied in federated learning systems for edge and fog environments, as well as in information systems of critical infrastructure facilities.