Game-Theoretic Optimized Federated Learning for Heterogeneous IoT Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Distributed Systems

Background:

• Federated object detection enables collaborative model training for IoT cameras without sharing raw data.
• Challenges include non-IID data, heterogeneous resources, unreliable networks, and participant selfishness.

Purpose of the Study:

• To propose GO-FedDet, a game-theoretic framework for optimized federated object detection in heterogeneous IoT applications.
• To address limitations of current federated learning approaches in real-world IoT scenarios.

Main Methods:

• Formulated client participation as a Stackelberg game and client selection as an exact-potential resource game.
• Embedded game-theoretic equilibria into a proximal detection objective.
• Developed a utility-aligned aggregation mechanism balancing contribution, cost, and fairness.

Main Results:

• Proved the existence of a Stackelberg equilibrium and convergence for the selection game.
• Derived a non-convex convergence bound considering client drift and partial participation.
• Demonstrated improved accuracy, reduced communication costs, and stabilized fairness in experiments.

Conclusions:

• GO-FedDet offers a robust solution for federated object detection in heterogeneous IoT environments.
• The game-theoretic approach effectively manages participant behavior and resource allocation.
• The framework shows significant improvements over existing methods in key performance metrics.