Efficient Wi-Fi Sensing for IoT Forensics with Lossy Compression of CSI Data

TL;DR

This paper addresses the challenge of storing and processing high-dimensional Wi-Fi CSI data in IoT-forensics by systematically evaluating lossy compression strategies. It compares traditional approaches (PCA with scalar and vector quantization) and deep learning-based methods (VAEs) across two forensic-relevant tasks: presence detection and activity recognition. Findings show that PCA+SQ can achieve dramatic data reductions with minimal sensing performance loss, while high-end DL methods can reach extreme compression but demand more resources and tuning; overall, PCA-based schemes offer the best practical balance for lightweight forensic deployments. The work demonstrates the feasibility of integrating lossy CSI compression into forensic Wi-Fi sensing pipelines, enabling long-term data retention and efficient storage without significantly compromising evidence quality.

Abstract

Wi-Fi sensing is an emerging technology that uses channel state information (CSI) from ambient Wi-Fi signals to monitor human activity without the need for dedicated sensors. Wi-Fi sensing does not only represent a pivotal technology in intelligent Internet of Things (IoT) systems, but it can also provide valuable insights in forensic investigations. However, the high dimensionality of CSI data presents major challenges for storage, transmission, and processing in resource-constrained IoT environments. In this paper, we investigate the impact of lossy compression on the accuracy of Wi-Fi sensing, evaluating both traditional techniques and a deep learning-based approach. Our results reveal that simple, interpretable techniques based on principal component analysis can significantly reduce the CSI data volume while preserving classification performance, making them highly suitable for lightweight IoT forensic scenarios. On the other hand, deep learning models exhibit higher potential in complex applications like activity recognition (achieving compression ratios up to 16000:1 with minimal impact on sensing performance) but require careful tuning and greater computational resources. By considering two different sensing applications, this work demonstrates the feasibility of integrating lossy compression schemes into Wi-Fi sensing pipelines to make intelligent IoT systems more efficient and improve the storage requirements in forensic applications.

Figures (8)

• Figure 1: Overview of the workflow proposed for efficient Wi-Fi sensing in forensics applications. Collected csi data is stored in a compressed form and retrieved after a potentially very long time. Lossy compression algorithms can be used to minimize the csi storage requirements if they are proven to retain good sensing accuracy.
• Figure 2: Methodology for evaluating the tradeoff between the compression ratio and sensing accuracy. In the compression stage, we consider several different algorithms; the sensing accuracy is evaluated in two target application scenarios, namely presence detection and activity recognition.
• Figure 3: Layout of the rooms in which the experiments are performed. In the presence detection case, the candidate enters/exits the room repeatedly, in the activity recognition case, the candidate performs several activities in the marked area.
• Figure 4: The value of $A^*$ over time (in blue). The red square wave indicates the ground truth when there is human presence (high value) or the room is empty (low value).
• Figure 5: Sketch of the pipeline for the dl-based classification approach when using the vae to compress the dataset.