Adversarial Contrastive Representation Learning for Passive WiFi Fingerprinting of Individuals
De Silva, Corentin Artaud · 2024
Extracting valuable patterns from a dataset involves isolating the pattern of relevance from any other influential attributes in the data generation process. This becomes important when trying to repurpose existing datasets that were collected for a specific objective. When repurposing datasets, irrelevant influential attributes cannot necessarily be treated as noise. This paper aims to develop a representation learning scheme to identify individual users based on their radio-frequency footprint (WiFi Fingerprinting) within a collection of WiFi access points in a building. The question pertains to whether a user-dependent signature can be isolated from the Received signal strength indicator (RSSI) across a network of WiFi access points, which can then be used to identify individuals in a building without the need for line of sight sensors such as cameras. We repurposed a popular dataset called UJIndoorLoc, which was collected to localize individual users based on the RSSI. We propose a novel adversarial contrastive representation learning scheme based on a triplet loss to extract user-dependent patterns while minimizing the influence of location-dependent patterns inherent within this dataset. The results indicate a robust user-dependent pattern that can be isolated with an accuracy of 0.87. Such coarse-grained identification methods based on RSSI can be combined with fine-grained Channel State Information (CSI) for passively identifying individuals without requiring cameras.