Prevention of Data Privacy and Recommendations using Federated Graph Neural Network

One type of distributed learning that makes use of edge devices for training is called federated learning. By keeping the real data on the users' devices and transmitting the learning parameters and gradient updates to the global server during the training process, it seeks to protect users' privacy. Instead of using user data directly, the global server uses these parameters for training, and the client’s devices can be used for local model tuning. In this thesis, we provide an overview of the learning paradigm and suggest a new federated recommender system framework that makes use of homomorphic encryption. However, federated learning is not without its drawbacks. While user privacy increases significantly, accuracy measurements experience a minor decline as a result. Furthermore, we demonstrate that using encrypted gradients for computations results in negligible impact on recommendation speed and guarantees a more secure method of transmitting user gradients to and from the global server. Our suggested work has been verified using two popular public datasets. In comparison to earlier research, the outputs of this study show a reduction in MAE and RMSE, supporting the efficacy and reasoning of FedGR (federated graph neural network for recommendation systems).