Evolutionary Algorithm-Based Ensemble of Aggregation Strategies in Federated Learning
IEEE Transactions on Evolutionary Computation, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/tevc.2026.3717202
- Dergi Adı: IEEE Transactions on Evolutionary Computation
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, BIOSIS, Compendex, INSPEC, Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
- Anahtar Kelimeler: Deep Learning, Ensemble Learning, Evolutionary Algorithm, Federated Learning, Transformer
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
Federated Learning (FL) has emerged as a valuable Machine Learning paradigm in various data-sensitive applications, such as occupancy detection and human activity recognition. One of the key elements to FL performance is the aggregation strategy, i.e., the way model parameters from clients are aggregated on the server before being updated. Several aggregation strategies have been proposed, each characterized by specific strengths and limitations. In this paper, we address the aggregation strategy selection problem by employing multiple aggregation strategies together, combining them in a complementary manner. Specifically, we consider several ensemble approaches, where up to six sub-global models, obtained with different baseline aggregation strategies, are maintained on the server. First, we consider coefficient-based ensembles, where the parameters of the sub-global models are weighted-averaged with either constant or adaptive coefficients, to create a single Ensemble Global Model (EGM). We then propose an ensemble approach based on an Evolutionary Algorithm (EA-EAFL), where instead aggregation occurs by selecting, for each layer of the EGM, among the corresponding layers in the sub-global models. This approach preserves the parameters of the baseline strategies at the layer level. We evaluate the proposed ensemble approaches on a Transformer, on six different human-related datasets, and a 2D Convolutional Neural Network, on two benchmark vision datasets, comparing them against each other, as well as against six different baseline aggregation strategies from the recent literature. Furthermore, we compare EA-EAFL against two recent EA-based FL approaches. Results reveal that the proposed approach consistently achieves superior accuracy, surpassing all other tested aggregation strategies.