Adaptive Weight Aggregation in Federated Learning for Brain Tumor Segmentation

Muhammad Irfan Khan, Mojtaba Jafaritadi, Esa Alhoniemi, Elina Kontio, Suleiman A. Khan

Abstract: We introduce similarity weighted aggregation, a principled and efficient method for regularized weight aggregation in federated learning. Our method is adapted to non-IID collaborators and is simultaneously cost-efficient. This is the first method to propose a sliding-window to select the collaborators, to the best of our knowledge. We demonstrate our method on the federate training task of the FeTS 2021 challenge. We proposed two variations coined Similarity Weighted Aggregation (SimAgg) and Regularized Aggregation (RegAgg). SimAgg results on internal validation data demonstrate that the proposed method outperforms the baseline FedAvg. The method SimAgg by our team HT-TUAS won 2nd position on both leaderboards in FeTS2021 challenge. SimAgg is the only method to be among the top performing methods on both the leaderboards, making it robust and reliable to data variations. Our solution is open sourced at: https://github.com/dskhanirfan/FeTS2021