SECURE DISTRIBUTED LEARNING: ROBUST DATA POISONING DETECTION FRAMEWORK USING INTELLIGENT MODEL INTEGRITY ANALYSIS

Abstract

Distributed Machine Learning (DML) has gained significant traction for its ability to train models collaboratively across multiple devices or servers without centralized data sharing. However, this distributed nature exposes DML systems to data poisoning attacks, where adversaries inject malicious or corrupted data to manipulate global model behavior. This paper proposes a robust data poisoning detection framework that combines intelligent integrity verification, gradient analysis, and anomaly detection techniques. The system identifies poisoned updates in federated environments by analyzing feature distribution shifts and inconsistent model gradients. Experimental validation demonstrates that the proposed method outperforms conventional statistical filters in both accuracy and resilience, effectively mitigating poisoning threats while preserving learning efficiency. This approach enhances trust and reliability in collaborative AI systems deployed across edge, cloud, and IoT infrastructures.