<ul><li>Federated learning (FL) for time series forecasting (TSF) can potentially lead to privacy risks due to gradient inversion attacks (GIA).</li><li>A study was conducted on inverting time series (TS) data across multiple TSF models and datasets, revealing unique challenges in reconstructing both observations and targets of TS data.</li><li>A novel GIA called TS-Inverse is proposed, which incorporates a gradient inversion model, unique loss function, and regularization techniques to improve the inversion of TS data.</li><li>TS-Inverse achieves significant improvement in sMAPE metric compared to existing GIA methods on TS data.</li></ul>

TS-Inverse: A Gradient Inversion Attack Tailored for Federated Time Series Forecasting Models

Discover more