Resumo

Machine learning optimizes performance in many embedded applications. A weak point of many learning solutions is the intensive use of data and computational resources required for training the model. By default, client devices send data to a solution developer's server to execute the training process in a more computationally powerful environment. However, this approach can compromise the client's privacy, as data is transmitted to third parties for processing. Federated learning solves this problem by training the model on the client devices, thus without sharing data. The trained models are then aggregated on the server to create a generalized version that can run on every client. The federated learning protocol involves selecting which clients will participate in each training round, with selection criteria focused on maximizing the number of clients per round, controlling fairness, lowering round discards, and managing resources. However, existing selection algorithms neglect the minimization of battery consumption, which is critical in scenarios where clients have limited resources. In this paper we propose a client selection mechanism for a federated learning protocol that considers energy, processing capacity, and network quality as determinant criteria for decision. Compared to a state-of-the-art selection technique, our algorithm saves resources while maintaining the model's accuracy.

Links