Federated Learning With Adaptive Regularization for Efficient Edge Data Corruption Detection in Edge Intelligence

Md. Rashedul Islam

doi:10.1109/TCC.2025.3623572

Federated Learning With Adaptive Regularization for Efficient Edge Data Corruption Detection in Edge Intelligence

2025-10-22

0

PRE

AI

DOI：10.1109/TCC.2025.3623572

Original

Original request for help

Save

Abstract

En 中文

Edge intelligence is an emerging distributed computing paradigm driven by the rapid proliferation of Internet of Things (IoT) devices and advancements in edge computing and artificial intelligence. With latency-sensitive data cached across multiple Edge Servers (ESs), efficient Edge Data Integrity Verification (EDIV) has become increasingly critical. Traditional ‘challenge-response’ EDIV methods incur high computation and communication costs by verifying all ESs indiscriminately, even when only a few may be corrupted. A recent Federated Learning (FL)-based framework partially mitigated this by identifying potentially corrupted ESs early using homogeneous activity data. However, under heterogeneous ES data, this approach suffers from reduced detection accuracy, slower convergence, and no clear guidance for subsequent verification rounds, thereby limiting overall cost reduction. To address these limitations, we propose Federated learning with Adaptive Regularizer-based Edge Data Integrity Verification (FedAR-EDIV), an FL-based framework equipped with an adaptive objective regularization strategy specifically designed to handle heterogeneous data distributions. FedAR-EDIV accurately identifies suspicious ESs during FL training, accelerates convergence, and significantly reduces computation and communication costs in the final EDIV stage. It achieves up to 16× faster communication and 9.1× lower computation cost compared to baseline EDIV methods, while maintaining detection accuracy above 99.78% on heterogeneous KDD99 activity data. Additionally, a dynamic reputation mechanism is applied after each EDIV round to reduce unnecessary audits on trustworthy ESs and allocate more scrutiny to potentially corrupted ones. Theoretical analysis verifies convergence, correctness, and security, while extensive experiments on heterogeneous datasets demonstrate superior accuracy, efficiency, and cost-effectiveness compared to existing methods.

Keywords:

Edge data integrity

edge computing

adaptive regularization

dynamic reputation

data integrity verification

Federated Learning With Adaptive Regularization for Efficient Edge Data Corruption Detection in Edge Intelligence

Abstract

Journal

Organization