The exponential emergence of smart connected objects, autonomous systems and embedded systems producing massive quantities of data in real time or in deferred time that require enormous storage capacity, backup and replication systems and optimal management of network bandwidth. In this article, we will discuss the different researches that have been conducted for the evolution of distributed, decentralized and secure federated learning architectures. Each of them applies a different data processing algorithm to optimize the training of a federated learning model in terms of data processing time, local training time and the computing and storage resources allocated to data processing by IoT device. We start with the edge federated learning architecture, in which we deploy a federated and distributed learning model while providing computing resources to smart devices to efficiently train a parallel and distributed learning model, the second clustered and secure federated learning architecture consists in involving in addition to servers and clients, gateways and IDS in the training phase, the last one aims to integrate the learning model into a blockchain system while applying a POW (Proof of Work) consensus mechanism that guarantees the integrity and confidentiality of the hyperparameters of the local learning models. All these architectures mentioned above have advantages and disadvantages that positively or negatively influence the federated learning phase in terms of the heterogeneity of data, models and resources embedded in autonomous systems and also on the secure exchange of hyperparameters of local models on the network. This consists in thinking about setting up a new secure and distributed architecture of federated learning based on new lightweight and secure network and telecommunication technologies.