Decentralized Federated Object Detection

With the increasing adoption of federated learning (FL) in privacy-sensitive applications, decentralized federated object detection (DFOD) has emerged as a promising approach to collaborative model training without a centralized authority. This thesis focuses on the implementation and optimization of an object detection model within a decentralized federated learning (DFL) framework. The primary objectives include (1) designing and deploying a federated object detection system on a decentralized platform, (2) enhancing its performance under non-IID data distributions through advanced aggregation techniques and adaptive learning strategies, and (3) evaluating its robustness against various poisoning attacks, such as label flipping, backdoor insertion, and model poisoning. The research aims to contribute to the security and efficiency of decentralized FL-based object detection systems.