Toward Design of a Scalable Federated Unlearning Framework for Trustworthy Edge Intelligence

Abstract

Federated learning (FL) enables collaborative model training across edge devices without centralizing raw data, but existing frameworks remain ill-equipped to support data privacy regulations mandated by GDPR, HIPAA, and CCPA. Once user data has influenced training, its verifiable removal becomes prohibitively expensive, particularly in non-IID and resource-constrained edge environments. This paper introduces a modular and scalable federated unlearning framework that unifies three complementary strategies: gradient subtraction, knowledge distillation, and checkpoint rollback, within an adaptive decision layer. A resource-aware checkpoint manager reduces storage costs through compression and pruning, while a privacy and trust layer integrates zero-knowledge proofs, differential privacy, and Merkle-based audit logs to provide verifiable guarantees of deletion. A non-IID-aware aggregator further preserves fairness across heterogeneous clients. Unlike prior approaches, our proposed framework systematically integrates rollback efficiency with formal privacy protections and auditability, offering a practical path toward trustworthy and regulation-compliant unlearning in domains such as healthcare, transportation, and smart agriculture.