DACFL: Dynamic Average Consensus-Based Federated Learning in Decentralized Sensors Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Federated Learning (FL) enables privacy-preserving model training on decentralized data from user devices.
• Centralized FL architectures face challenges like communication congestion and single points of failure in sensor networks.

Purpose of the Study:

• To propose a decentralized Federated Learning approach, Dynamic Average Consensus-based Federated Learning (DACFL), for sensor networks.
• To address the limitations of centralized FL, particularly model aggregation in decentralized settings.

Main Methods:

• Transformed FL model aggregation into a dynamic average consensus problem, treating local training as a discrete-time series.
• Employed first-order dynamic average consensus (FODAC) for decentralized model aggregation and consistency.
• Incorporated neighbor model averaging for initialization to mitigate local over-fitting with non-i.i.d data.

Main Results:

• DACFL demonstrated feasibility in both time-invariant and time-varying network topologies.
• Achieved significant average accuracy increases compared to CDSGD and D-PSGD on Fashion-MNIST dataset.
• With i.i.d data, accuracy increased by 19-34% (vs. CDSGD) and 3-10% (vs. D-PSGD).
• With non-i.i.d data, accuracy increased by 30-50% (vs. CDSGD) and 0-10% (vs. D-PSGD).

Conclusions:

• DACFL offers a robust and efficient decentralized solution for Federated Learning in sensor networks.
• The method effectively handles model aggregation and improves performance, especially under non-i.i.d data distributions.