MFedBN: Tackling Data Heterogeneity with Gradient-Based Aggregation and Advanced Distribution Skew Modeling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Federated Learning (FL) facilitates collaborative training on edge devices, preserving data privacy.
• FL performance significantly declines with non-Independent and Identically Distributed (non-IID) data, hindering real-world applications.
• Existing aggregation strategies struggle to maintain performance under diverse non-IID data conditions.

Purpose of the Study:

• To evaluate aggregation strategies in non-IID FL environments.
• To propose novel methods for generating skewed datasets representing various non-IID types.
• To introduce and validate a new aggregation strategy, MFedBN, for improved FL performance.

Main Methods:

• Developed data generation techniques for Feature Distribution Skew, Label Distribution Skew, Same Label/Different Features skew, and Same Features/Different Label skew.
• Introduced Modified Federated via Local Batch Normalization (MFedBN) employing server-side gradient-style updates with varied learning rates.
• Conducted experiments on Commercial Vehicles Sensor and NF-UNSW-NB15 datasets.

Main Results:

• MFedBN outperformed the baseline FedBN in most scenarios.
• Achieved high test accuracies: up to 85% on Commercial Vehicles Sensor and 99.98% on NF-UNSW-NB15.
• Demonstrated improved convergence stability and generalization for MFedBN in heterogeneous FL environments.

Conclusions:

• MFedBN effectively enhances FL performance and robustness across various non-IID data skews.
• The proposed data generation methods provide a platform for evaluating FL strategies in non-IID settings.
• This work advances the applicability of privacy-preserving FL in real-world IoT monitoring and network intrusion detection.