Una piattaforma sperimentale che consenta l'apprendimento distribuito su set di dati privati

Federated Learning (FL) is a collaborative, distributed machine learning (ML) approach that re- cently gained research interest due to its privacy-preserving properties.[2] The basic concept behind FL is moving ML models trained on different datasets instead of moving data itself, avoiding most of the issues implied by this latter process. On the other hand, function as a service (FaaS) is a cloud computing paradigm that enables clients to develop and deploy applications to be executed on demand. Billing of such functionalities is thus calculated on the actual execution time and resource consumption of the service. Building on the properties of these two techniques, we propose the FL as a service (FLaaS) platform. This approach allows decentralized learning workloads on local, private datasets, the access to which is granted as a service. ML practitioners can thus develop an ML model, select a series of datasets available on the FLaaS infrastructure and automatically deploy a secure, privacy-preserving FL training on the specified hosts, exploiting both their data and computational power. This master thesis proposes an experimental implementation of the FLaaS platform, exploiting many state-of-the-art software products, such as Kubernetes, Docker, RabbitMQ, and ONNX, to guarantee high levels of performance and security. FLaaS results in an automated infrastructure facilitating the distributed deployment, execution, and management of containerized FL workloads, allowing users to configure the process to suit their needs at best. We finally test the FLaaS platform under different conditions, varying the computational resources available to the system (number of computing nodes) and the FL task hyperparameters (model, dataset), exploiting the FLOWER FL framework and the MNIST dataset as the experimental backbones.