A Federated Learning Framework for Privacy-Preserving Cyber Threat Detection

Abstract

The escalating frequency and sophistication of cyberattacks has driven growing interest in machine-learning-based intrusion and threat detection systems that learn from network and endpoint telemetry across many organizations. Centralized training pipelines, however, require member organizations to pool raw security logs into a shared repository, an arrangement that is frequently prohibited by data-protection regulation, competitive concerns, and the sensitivity of breach-related information. This paper presents FedShield, a federated learning framework that enables geographically distributed Security Operations Centers (SOCs) to collaboratively train a shared threat-detection model while keeping raw traffic and log data confined to each participant's own infrastructure. FedShield combines a hybrid convolutional-recurrent local detector, a secure-aggregation protocol built on pairwise masking, and a differential-privacy mechanism applied to client updates to bind the information that any single gradient contribution can reveal about a participating organization's network. A gradient-similarity filter is further used to down-weight or exclude poisoned or anomalous client updates before aggregation, mitigating model-poisoning and label-flipping attacks that federated settings are known to be vulnerable to. FedShield was evaluated in a simulated consortium of up to 100 client organizations using partitions of the CIC-IDS2018 and N-BaIoT traffic datasets under non-independent and identically distributed (non-IID) data conditions. The framework reached 97.3% detection accuracy with a 2.6% false-positive rate after 40 communication rounds, within 1.1 percentage points of a centralized baseline trained on pooled raw data, while local-only client models trained in isolation plateaued at 88.2% accuracy. Under a differential-privacy budget of epsilon equal to 4, FedShield retained 94.1% accuracy, illustrating a favorable privacy-utility trade-off. The aggregation was also poison resistant, making successful label flipping attacks less than 4% compared to 41% success when using the original aggregation. These findings suggest that federated learning, with secure aggregation and calibrated differential privacy, provides a feasible approach to sharing cross-organizational threat intelligence that respects data sovereignty and regulatory requirements while maintaining detection performance comparable to centralized approaches.