Synthetic Model Aggregation Strategies in Federated Learning: Cloning and Crossover-Based Approaches

Federated Learning (FL) enables the collaborative training of machine learning models without requiring the sharing of local datasets. However, the performance of FL systems can degrade when the number of participating clients decreases or when data distributions are highly heterogeneous. This paper investigates alternative aggregation strategies based on synthetic models generated directly from existing trained models, rather than from synthetic training data. In particular, we analyze two approaches: model cloning and crossover-based model generation inspired by genetic algorithms. Both strategies aim to improve robustness when only a subset of clients can participate in the standard training process. Experimental results show that cloning maintains stable performance across different participation levels, while crossover-based aggregation introduces greater variability but may enhance model diversity. Overall, the results highlight both the potential and the limitations of synthetic model aggregation in decentralized learning environments.