Author(s)

Dr MOHAMED AZHARUDHEEN A, Mrs D Lalitha, Mr A Akilan

  • ISSN (P): 3139-8464
  • Manuscript ID: 141309
  • Volume: 2
  • Issue: 11
  • Pages: 314–338

Subject Area: Computer Science and IT

DOI: https://doi.org/10.64643/jatir.141309
Abstract

The rapid growth of distributed big data environments across healthcare, financial services, Internet of Things (IoT), and cloud computing has increased the demand for secure and privacy-preserving machine learning mechanisms. Federated Learning (FL) enables collaborative model training across distributed data sources without transferring raw data to a centralized server. However, conventional federated learning approaches remain vulnerable to privacy leakage through model updates, membership inference attacks, gradient inversion, and malicious client participation. Furthermore, existing privacy-preserving mechanisms often introduce computational overhead and reduce analytical utility, while conventional aggregation strategies may fail to mitigate Byzantine and model poisoning attacks. To address these limitations, this research proposes a Trust-Aware Byzantine-Resilient Federated Learning with Adaptive Differential Privacy (TABR-FL-DP) framework for secure big data analytics. The proposed architecture integrates adaptive deep feature extraction, client-level differential privacy, trust-based client evaluation, Byzantine-resilient model aggregation, and fault-tolerant federated optimization. Initially, distributed clients perform local data preprocessing and feature extraction without exposing sensitive information. Gradient clipping and calibrated Gaussian noise are subsequently applied to protect individual client contributions under a formally accounted differential privacy mechanism. A trust evaluation engine assesses client reliability based on update consistency, historical behavior, and abnormal model-update characteristics. The proposed robust aggregation mechanism combines trust-weighted contributions with Byzantine-resilient outlier mitigation to reduce the influence of malicious participants. Additionally, an adaptive privacy-budget optimization mechanism balances privacy protection, analytical utility, communication overhead, and model convergence. The proposed framework will be evaluated using benchmark cybersecurity datasets, including CIC-IDS2017 and UNSW-NB15, against conventional centralized learning, FedAvg, and established privacy-preserving and Byzantine-resilient federated learning approaches. Performance will be assessed through classification accuracy, precision, recall, F1-score, differential privacy guarantees, attack success rate, communication overhead, latency, and fault tolerance. The research aims to establish a reliable and scalable privacy-preserving federated learning framework for secure and high-availability big data analytics.

Keywords
Federated LearningDifferential PrivacyByzantine-Resilient Aggregation; Trust-Aware Learning; Big Data Security; Model Poisoning Detection; Privacy-Preserving Machine Learning; Adaptive Privacy Budget; Deep Learning; Intrusion Detection.