Federated Learning for Fake News Detection and Data Privacy Preservation

People are overwhelmed by a huge amount of news they receive every day. Many times, they may relay the news to their family members or friends if they deem the news is relevant to the family members or friends, but not all the news they receive is true. According to a study, 38% of Americans have the experience of sharing fake news with others, and people usually trust the news more if it is from the persons they know. Various machine learning methods have been used to identify fake news, and each has its pros and cons. This research builds a simple, effective fake-news detection system by using sentiment and emotion analyses. Instead of focusing on improving the detection accuracy, this research emphasizes on preserving data privacy while training the model by using the method of federated learning (FL). After each client in FL trains its own model by using its local data, the parameter weights of all clients are collaboratively sent to the server and aggregated to train its global model. Experiment results show the proposed federated-learning method successfully builds a fake-news detection system and preserves the data privacy at the same time.