Network-aware federated neural architecture search

Deep Learning (DL) has significantly transformed various aspects of daily life by empowering technological innovations. The cooperation between DL and the Internet of Things (IoT) has further advanced these developments, allowing smart devices to serve as both data sources and endpoints for DL-powered applications. However, the success of DL relies on optimal Deep Neural Network (DNN) architectures, and manually developing such systems requires extensive expertise and time. Automated machine learning methods, such as Neural Architecture Search (NAS), have emerged to automate the search for the best-performing neural architectures. Meanwhile, Federated Learning (FL) addresses data privacy concerns by enabling collaborative model development without exchanging the private data of clients. This thesis focuses on the collaboration of Federated Learning and Neural Architecture Search, addressing challenges exposed by IoT devices with limited resources and changing network conditions. Contributions include the development of an open-source federated neural network pruning framework with network emulation support, aiming to simulate real-world scenarios and efficiently reduce the size of DNNs in a distributed network. A novel client grouping algorithm is proposed to optimize communication efficiency, considering both varying class distributions and network bandwidth conditions. Moreover, a detailed exploration of the impact of network bandwidth on federated learning, especially communication rounds, provides valuable insights for designing robust federated learning systems. These contributions collectively enhance the understanding of federated neural architecture search where computational efficiency, data privacy, and network conditions show significant roles.