Enhancing Privacy in IoT-Enabled Digital Infrastructure: Evaluating Federated Learning for Intrusion and Fraud Detection
Amogh Deshmukh, Josep Lluís de la Rosa, Raul Villamarin Rodriguez, Sandeep Dasari · Sensors · 2025
Challenges in implementing machine learning (ML) include expanding data resources within the finance sector. Banking data with significant financial implications are highly confidential. Diverse breaches and privacy violations can result from a combination of user information from different institutions for banking purposes. To address these issues, federated learning (FL) using a flower framework is utilized to protect the privacy of individual organizations while still collaborating through separate models to create a unified global model. However, joint training on datasets with diverse distributions can lead to suboptimal learning and additional privacy concerns. To mitigate this, established FL algorithms such as federated averaging (FedAvg), federated proximal (FedProx), and federated optimization (FedOpt), previously introduced in the literature, are evaluated in our study. These methods work with data locality during training at local clients without exposing data, while maintaining global convergence to enhance the privacy of local models within the framework. In this analysis, the UNSW-NB15 and credit datasets were employed, utilizing precision, recall, accuracy, F1-score, ROC, and AUC as performance indicators to demonstrate the effectiveness of applying FedAvg, FedProx, and FedOpt within the flower framework. The algorithms considered were subjected to an empirical study, which revealed significant performance benefits when using the flower framework. Consequently experiments were conducted over 50 rounds using the UNSW-NB15 dataset, which achieved accuracies of 99.87% for both FedAvg and FedProx and 99.94% for FedOpt. Similarly, with the credit dataset under the same conditions, FedAvg and FedProx achieved accuracies of 99.95% and 99.94%, respectively, while FedOpt also achieved comparable accuracy. These simulation-based results suggest that the proposed framework is effective on widely used benchmark datasets and, with further validation in deployed environments, has the potential to support secure and privacy-preserving collaborative machine learning across various domains.