Federated Learning is a method of training a Machine Learning or Deep learning model using individual dataset from different users without the need of sharing the data. Here the data of individual users are training on their local system and the learning is shared among them to get a global optimal model. In this work a Neural Machine Translation work on a resource poor language, i.e. English-Assamese language pair is performed using the concept of Federated Learning. Two different Federated Optimization Technique, i.e., FedOpt and FedYogi are used for performing the Federated Aggregation of all the individual client models. Two different Federated Learning architectures consisting of different number of clients (i.e., 4 and 8) are trained on a simulated environment of a single system. The Neural Machine Translation model is built based on Transformer encoder decoder model. The result were evaluated based on Bilingual Evaluation Understudy (BLEU) and CHaRacter-level F-score (CHRF) score.