Privacy Leakage in Federated Home Applications Using Gradient Inversion Algorithms

With advances in smart metering infrastructure, household electricity metering data are remotely collected, leading to concerns about household privacy leakage. Federated learning is a promising solution because it avoids direct data uploading. However, recent research shows that the gradient of federated learning contains a certain amount of private information that can be recovered using gradient inversion algorithms. This paper proposes an explainable algorithm to locate the key factors affecting privacy leakage and vulnerable data in home application scenarios. Simulations show comparative results of privacy leakages in different situations and reveal that for home Artificial Intelligence applications, smaller batch sizes, training iterations, and extreme values are prone to causing privacy leaks. Based on that, the advice for protecting federated learning privacy under gradient inversion algorithms is summarized.