Federated Learning and Privacy-Preserving AI: Emerging Approaches for Distributed Intelligent Computing — A Systematic Review of Algorithms, Privacy Mechanisms, and Deployment Trade-offs
Abstract
ABSTRACT
The growth of sensitive data held by hospitals, banks, mobile devices, and industrial edge nodes has created a basic conflict between the demand for data-driven intelligence and the legal and ethical duty to protect individual privacy. Federated learning (FL) addresses this conflict by moving model training to the data instead of moving data to a central server, yet FL alone does not guarantee privacy, because shared gradients and model updates can leak training information. This review-based study synthesises 35 peer-reviewed papers, preprints, books, and regulatory documents published between 2009 and 2023, following a protocol adapted from systematic review guidelines for software engineering. Each source was analysed along five dimensions: statistical heterogeneity, communication cost, privacy threat model, privacy mechanism, and deployment domain. The reported evidence shows that federated averaging reduces communication rounds by one to two orders of magnitude; that label-skewed non-IID data can reduce accuracy by up to 51% on CIFAR-10; that differentially private training retains roughly 97% accuracy on MNIST at a privacy budget of ? = 8 but falls to about 90% at ? = 0.5; and that secure aggregation protects individual updates but not the aggregate model. A qualitative comparison of five privacy mechanisms is presented, and a layered defence-in-depth framework is proposed for selecting mechanisms in cross-device and cross-silo deployments.
KEYWORDS: Federated learning, Privacy-preserving AI, Differential privacy, Secure aggregation, Homomorphic encryption, Non-IID data, Edge intelligence, Distributed computing, Gradient leakage, Data protection
The growth of sensitive data held by hospitals, banks, mobile devices, and industrial edge nodes has created a basic conflict between the demand for data-driven intelligence and the legal and ethical duty to protect individual privacy. Federated learning (FL) addresses this conflict by moving model training to the data instead of moving data to a central server, yet FL alone does not guarantee privacy, because shared gradients and model updates can leak training information. This review-based study synthesises 35 peer-reviewed papers, preprints, books, and regulatory documents published between 2009 and 2023, following a protocol adapted from systematic review guidelines for software engineering. Each source was analysed along five dimensions: statistical heterogeneity, communication cost, privacy threat model, privacy mechanism, and deployment domain. The reported evidence shows that federated averaging reduces communication rounds by one to two orders of magnitude; that label-skewed non-IID data can reduce accuracy by up to 51% on CIFAR-10; that differentially private training retains roughly 97% accuracy on MNIST at a privacy budget of ? = 8 but falls to about 90% at ? = 0.5; and that secure aggregation protects individual updates but not the aggregate model. A qualitative comparison of five privacy mechanisms is presented, and a layered defence-in-depth framework is proposed for selecting mechanisms in cross-device and cross-silo deployments.
KEYWORDS: Federated learning, Privacy-preserving AI, Differential privacy, Secure aggregation, Homomorphic encryption, Non-IID data, Edge intelligence, Distributed computing, Gradient leakage, Data protection