One-liner: Exploring how lossy compression of Wi-Fi Channel State Information (CSI) can make IoT forensics more efficient — cutting storage needs without losing key sensing accuracy.
This project looks at how Wi-Fi sensing — which uses CSI data from regular Wi-Fi signals — can detect human activity without extra sensors. It’s a growing area in intelligent IoT systems and can also support forensic investigations where sensor data is limited or unavailable.
A key issue with CSI-based sensing is that the data is huge. High-dimensional CSI samples are expensive to store, transmit, and process, especially on low-power IoT devices.
Our work studies how lossy compression can help, and how much compression you can apply before sensing accuracy starts to drop. We compare traditional methods with a deep learning–based approach.
We evaluated four lossy compression techniques for CSI data. Each one offers a different balance between storage efficiency and accuracy:
-
PCA (Principal Component Analysis)
Projects CSI frames onto a smaller set of orthogonal components. It’s fast, easy to interpret, and works well for sensing tasks. -
Scalar Quantization (SQ)
Quantizes CSI values or PCA coefficients into a limited number of levels. The number of bits per sample directly controls compression and quality. -
Vector Quantization (VQ)
Groups CSI vectors using clustering (like k-means) and represents each window by its cluster index. It captures relationships between subcarriers better than SQ. -
Variational Autoencoder (VAE)
Uses a neural network to learn a compact latent representation of the CSI data. It can achieve very high compression ratios, though it needs more compute and tuning.
- Preprocessing: Extract and normalize CSI amplitude; remove noisy or irrelevant subcarriers.
- Windowing: Segment data into fixed-length windows (e.g., 3 s for 'presence detection', longer for 'activity recognition').
- Compression: Apply PCA, SQ, VQ, or VAE to each window and store the compressed version.
- Reconstruction & Evaluation: Decompress the data when needed and run it through classifiers to measure sensing performance.
- CSI data can be compressed by several orders of magnitude with little to no loss in sensing accuracy.
- Deep learning–based compression (VAE) performs best on complex tasks like activity recognition, achieving up to 16,000× compression while maintaining accuracy.
- Traditional methods like PCA or VQ are still valuable for lightweight setups that need quick, interpretable results.
Overall, integrating lossy compression into Wi-Fi sensing pipelines makes IoT forensics more scalable and practical, especially when working with limited bandwidth and storage.
This repository includes the code and notebooks used in the experiments described in the paper.
Note: Datasets are not included.