The traditional cloud-centric approach for Deep Learning (DL) requires\ntraining data to be collected and processed at a central server which is often\nchallenging in privacy-sensitive domains like healthcare. Towards this, a new\nlearning paradigm called Federated Learning (FL) has been proposed that brings\nthe potential of DL to these domains while addressing privacy and data\nownership issues. FL enables remote clients to learn a shared ML model while\nkeeping the data local. However, conventional FL systems face several\nchallenges such as scalability, complex infrastructure management, and wasted\ncompute and incurred costs due to idle clients. These challenges of FL systems\nclosely align with the core problems that serverless computing and\nFunction-as-a-Service (FaaS) platforms aim to solve. These include rapid\nscalability, no infrastructure management, automatic scaling to zero for idle\nclients, and a pay-per-use billing model. To this end, we present a novel\nsystem and framework for serverless FL, called FedLess. Our system supports\nmultiple commercial and self-hosted FaaS providers and can be deployed in the\ncloud, on-premise in institutional data centers, and on edge devices. To the\nbest of our knowledge, we are the first to enable FL across a large fabric of\nheterogeneous FaaS providers while providing important features like security\nand Differential Privacy. We demonstrate with comprehensive experiments that\nthe successful training of DNNs for different tasks across up to 200 client\nfunctions and more is easily possible using our system. Furthermore, we\ndemonstrate the practical viability of our methodology by comparing it against\na traditional FL system and show that it can be cheaper and more\nresource-efficient.\n