Preserving Privacy in CSI-based Human Activity Recognition: A Data Obfuscation Case Study

Human Activity Recognition (HAR) techniques play a key role in identifying and categorizing human activities based on environmental information. More recently, the use of Channel State Information (CSI) has gained momentum because this information can be extracted in a non-intrusive manner. CSI-based algorithms leverage the correlation between CSI dynamics of wireless transmissions and human body movements. However, privacy concerns may arise, as this approach may inadvertently disclose sensitive information about individuals' movements, habits, and behaviors. In this context, this study investigates the challenge of preserving user privacy in CSI-based HAR for eHealth Ambient Assisted Living (AAL) applications. More specifically, the impact of simple filter-based CSI obfuscation is evaluated on the accuracy performance of a HAR model that makes use of a Long-Short-Term Memory (LSTM) algorithm. Using a publicly available dataset, the accuracy between the original and the obfuscated versions of the dataset generated by simple filtering techniques is compared. The results show significant performance degradation when data is obfuscated, with a HAR accuracy degradation compared to the original results ranging from a minimum of 17.9% to a maximum of nearly 90%. Such results prove that, even with simple obfuscation techniques, the privacy of CSI-enabled HAR-based systems can be sufficiently preserved.