Types of Federated Learning and Aggregation Techniques
S Shailesh, Joseph James · Apple Academic Press eBooks · 2024
The contemporary advancements in the areas ofmachine learning and artificial intelligence (AI) open up many opportunities to adopt these technologies in real life. However, concerns also arise when dealing with confidential user data. In traditional machine learning models, the distributed learning models themselves fail to maintain the privacy of the autonomous data from each client. They collect all the data to a single point and perform the learning process at the central server. Researchers in this field have come up with a promising solution called Federated Learning, which ensures the privacy of each client’s data. The federated learning approach is capable of training machine learning models for each client independently and combining those independent models to build a general global model. This chapter explains different variants or types offederated learning approaches such as ‘Horizontal Federated Learning,’ Vertical Federated Learning,’ and ‘Federated Transfer Learning’. Horizontal federated learning utilizes different features of the same sample set, while in vertical federated learning, all the clients utilize different sets of samples with a common set of features. In federated transfer learning, the clients have the freedom to use any sample set and any feature 24 set. This chapter also explains different aggregation techniques used in federated learning architectures. The idea of aggregation is to collect the model parameters from all autonomous clients and combine them to create a single global model. This task requires more technical soundness as it is the most critical phase in federated learning. The three specific methods: One Model Selection (OMS), All Model Averaging (AMA), and Best Model Averaging (BMA) are elaborated with detailed algorithms and use cases. Along with the classical aggregation techniques, the article explains three secure aggregation protocols: Homomorphic encryption-based aggregation, Blockchain-based aggregation, and TEE-based aggregation. All the techniques are explained in the context of homogeneous federated learning architecture, which uses the same learning algorithm specifications in all client endpoints.