A Game Theoretic Adversarial Synthetic Data Generation Method to Address Privacy Concerns in the Use of Deep Learning Models for IoT Applications

Internet of Things (IoT) applications are widely prevalent in the age of Industry 4.0. These applications generate vast amounts of data that need to be processed efficiently. In some of these applications, there may be privacy concerns associated with the generated data being used to train Artificial Intelligence (AI) models. Thus, solutions are needed that can address such problems. This paper develops a methodology to generate diversified synthetic datapoints for IoT datasets, using adversarial machine learning in a game-theoretic setting. The proposed method jointly optimizes two 2-player games being played simultaneously, where the players in each 2-player game have a competing objective. The experimental results on a publicly-available dataset demonstrate that when Machine Learning (ML) models are retrained using only these synthetic datapoints, they perform better than the ML models trained on the original smart meter data, thus alleviating the need for using private data for training ML models.