Learning Private Neural Language Modeling with Attentive Aggregation (1812.07108v2)

Abstract: Mobile keyboard suggestion is typically regarded as a word-level LLMing problem. Centralized machine learning technique requires massive user data collected to train on, which may impose privacy concerns for sensitive personal typing data of users. Federated learning (FL) provides a promising approach to learning private LLMing for intelligent personalized keyboard suggestion by training models in distributed clients rather than training in a central server. To obtain a global model for prediction, existing FL algorithms simply average the client models and ignore the importance of each client during model aggregation. Furthermore, there is no optimization for learning a well-generalized global model on the central server. To solve these problems, we propose a novel model aggregation with the attention mechanism considering the contribution of clients models to the global model, together with an optimization technique during server aggregation. Our proposed attentive aggregation method minimizes the weighted distance between the server model and client models through iterative parameters updating while attends the distance between the server model and client models. Through experiments on two popular LLMing datasets and a social media dataset, our proposed method outperforms its counterparts in terms of perplexity and communication cost in most settings of comparison.