PRIVACY-PRESERVING FEDERATED LEARNING FRAMEWORK USING DIFFERENTIAL PRIVACY, BYZANTINE-RESILIENT AGGREGATION, AND CARBON-AWARE DISTRIBUTED TRAINING FOR SECURE, FAIR, AND SUSTAINABLE ARTIFICIAL INTELLIGENCE
Keywords:
federated learning, differential privacy, Byzantine resilience, carbon-aware training, secure aggregation, fairness, distributed learning, responsible AI.Abstract
Federated learning has emerged as a foundational paradigm for training machine learning models across distributed data silos without requiring raw data centralization. However, its deployment at scale remains constrained by three interconnected challenges: the privacy–utility trade-off introduced by differential privacy, the security–performance tension associated with Byzantine-resilient aggregation, and the accuracy–sustainability conflict arising from the communication and energy demands of iterative distributed training. This study presents a Privacy-Preserving Federated Learning Framework that jointly addresses these challenges through adaptive Rényi differential privacy with dynamic noise calibration, multi-metric Byzantine-resilient aggregation combining cosine-similarity filtering with coordinate-wise robust aggregation, carbon-aware participant scheduling with gradient compression, and fairness-aware aggregation correction.
The framework was evaluated across three benchmark datasets—CIFAR-10, MIMIC-III, and AGNews—under heterogeneous non-IID data distributions, multiple privacy budgets, four Byzantine attack strategies, and Byzantine participant rates of up to 40%. At ε = 8 and a 30% Byzantine participant rate, the framework achieved 82.1% accuracy on CIFAR-10 and 80.8% on MIMIC-III, substantially outperforming standard FedAvg under equivalent attack conditions. The multi-metric aggregation mechanism provided a 34.2% accuracy recovery over FedAvg under the most challenging Byzantine attack scenario. Adaptive privacy accounting reduced cumulative privacy-budget consumption by 28.4% compared with fixed-noise approaches while maintaining equivalent formal privacy guarantees. Carbon-aware scheduling and gradient compression reduced the overall training carbon footprint by 41.6%, while fairness-aware aggregation-maintained subgroup performance disparities below a Gini coefficient of 0.08. Overall, the results demonstrate that privacy, Byzantine resilience, predictive utility, fairness, and environmental sustainability can be jointly optimized within a unified federated learning architecture, providing a practical foundation for secure, fair, and sustainable distributed AI.
