MockiFi: CSI Imitation using Context-Aware Conditional Neural Process for Zero-shot Learning

WiFi sensing technology has emerged as a promising technology for activity recognition, leveraging the Channel State Information (CSI) to capture fine-grained movement details. However, the difficulty and scarcity in collecting the required training CSI data with variations hinder the development and deployment of practical WiFi sensing systems in different settings. This paper presents MockiFi, a novel system that learns WiFi CSI data transformations across different activities of known individuals and generates CSI data for the activities of new individuals using their base activity CSI data and by mimicking transformations learned from known individuals. Our approach employs a Conditional Neural Process (CNP) to synthesize realistic activity patterns through the learning of pattern transitions using a cosine similarity-based loss function. The effectiveness of our approach is validated through extensive experiments, achieving a high cosine similarity score between the generated and real activity data, indicating the precision and reliability of the generated action fingerprints. We also show that a classifier trained on synthetic data of a new person can successfully recognize their actual activities, demonstrating zero-shot learning capabilities. These results show that MockiFi can help develop customized WiFi sensing systems without the need for collecting excessive new training data, and thus can facilitate their practical usage.