Federated Machine Learning for Privacy-Preserving Predictive Analytics and Distributed Decision Intelligence

In distributed and heterogeneous computing environments, federated machine learning has proven to be a viable approach to privacy-preserving predictive analytics. This study introduced a federated system for anomaly detection to aid in collaborative learning and distributed decision intelligence, without aggregating raw network data on the central server. An unsupervised federated autoencoder was applied to learn traffic representations across decentralized clients, and Federated Averaging was applied to collect locally trained model parameters into a global model. To assess the model performance, reconstruction loss, accuracy, precision, recall, specificity, F1 score, and ROC AUC were calculated. The federated model shows that the mean reconstruction loss gradually decreases with the number of communication rounds, indicating convergence, and the loss in the 10th communication round is 0.6754, which is approximately a 32.14% decrease from the loss in the 1st communication round. With respect to recognition of benign traffic, the model attained accuracy rate of 70.44% and specificity of 91.96% which is good. The attack recall was however, limited to 6.36%, F1-score of 9.76% and ROC-AUC of 0.365, which showed that the system discriminated poorly between malicious and non-malicious observations. The results indicate that convergence from a federated perspective does not always result in intrusion detection success. The study also highlights the importance of adaptive aggregation.