Inferring Personal Attributes with a Mmwave Radar

Various devices can be utilised to sense people in smart environments, and in this task, millimetre-wave (mmWave) radars have recently gained attention. This paper presents a novel investigation of the risk posed by mmWave radars that could be used in Internet of Things (IoT) scenarios, focusing on the potential for an attacker to access personal infor-mation without the user's consent. We introduce INFORMER: INFerring persOnal attRibutes with a MmwavE Radar, a new perspective on collecting information from personal features using gait. INFORMER utilises point clouds to capture the gait of volunteers. The mean information from the point clouds is then analysed to infer information from five attributes: height, weight, gender, waist, and arm span. Three variants of IN-FORMER scrutinised these sequences: one Deep Learning (DL) model and two Machine Learning (ML) models. The DL model showed accuracy up to 82%, demonstrating superior height and weight inference performance, while gender inference showed comparatively lower accuracy. Additionally, the first model ML, Singular Value Decomposition (SVD), tends to perform better in some attributes, such as waist, but worse attributes than the other models in features, such as weight. Lastly, the Principal Component Analysis (PCA) presents comparable results. Specifically, it shows the best performance in gender inference.