about | r&d | publications | courses | people | links

A. Sotiropoulou, C. Troussas, A. Krouska, Ph. Mylonas, C. Sgouropoulou, I. Voyiatzis

FedBigAnalytics: Privacy-Preserving Federated Learning for Scalable Big Data Analytics with Adaptive Aggregation and Differential Privacy

6th International Conference on Novel & Intelligent Digital Systems (NIDS 2026), September 23-25, 2026, Athens, Greece

ABSTRACT

As data across distributed sources are growing exponentially, a fundamental tension has been created between the analytical power of centralised big data platforms and the privacy, sovereignty, and regulatory constraints that prohibit data centralisation. Federated learning offers a principled resolution by training models across decentralised data silos without exchanging raw data, but exist-ing federated approaches suffer from statistical heterogeneity across nodes, communication inefficiency at scale, and insufficient formal privacy guaran-tees. This paper proposes FedBigAnalytics, a novel framework that integrates three innovations to address these challenges simultaneously: (1) an adaptive aggregation mechanism that dynamically weights local model updates based on data quality, distribution divergence, and node reliability, outperforming stand-ard FedAvg and FedProx aggregation; (2) a communication-efficient gradient compression scheme that reduces bandwidth requirements by 73% with less than 0.4% accuracy loss; and (3) a calibrated differential privacy module that provides formal (ε, δ)-privacy guarantees while maintaining 96.3% of non-private accuracy. We evaluate FedBigAnalytics on three large-scale real-world datasets (healthcare records across 12 hospitals, financial transaction logs from 8 institutions, and IoT sensor streams from 32 edge nodes) totalling 47 million records. Our framework achieves 91.8% accuracy, outperforming FedAvg by 7.1 percentage points and FedProx by 5.5 points, while converging 2.4× faster and scaling sub-linearly to 64 nodes. Under a privacy budget of ε = 1.0, Fed-BigAnalytics maintains 89.4% accuracy versus 81.3% for standard differential-ly private FedAvg. These results demonstrate that privacy-preserving federated analytics can match or exceed centralised performance on heterogeneous big data workloads

23 September, 2026

A. Sotiropoulou, C. Troussas, A. Krouska, Ph. Mylonas, C. Sgouropoulou, I. Voyiatzis, "FedBigAnalytics: Privacy-Preserving Federated Learning for Scalable Big Data Analytics with Adaptive Aggregation and Differential Privacy", 6th International Conference on Novel & Intelligent Digital Systems (NIDS 2026), September 23-25, 2026, Athens, Greece

[ BibTex] [ Print] [ Back]

© 00 The Image, Video and Multimedia Systems Laboratory - v1.12