PaperlayerView exact revision on arXiv

Source-linked AI summary

SenseFi: A Library and Benchmark on Deep-Learning-Empowered WiFi Human Sensing

Jianfei Yang, Xinyan Chen, Dazhuo Wang, Han Zou, Chris Xiaoxuan Lu, Sumei Sun, Lihua Xie

arXiv:2207.07859v3cs.LGcs.AIeess.SP

TL;DR

WiFi sensing lacks a sufficiently comprehensive public benchmark despite rapid progress in deep-learning-based applications. The paper introduces SenseFi, an open benchmark and model-zoo library that evaluates models and learning schemes across CSI datasets, platforms, and tasks, with experiments yielding practical guidance for WiFi sensing research and applications.

  • Problem

    Deep-learning WiFi sensing studies mainly target individual tasks, while realistic applications face expensive annotation and environmental distribution shifts.

  • Method

    SenseFi reviews deep models, benchmarks them on raw and processed CSI data across WiFi platforms, and evaluates transfer learning and unsupervised learning schemes.

  • Results

    The experiments compare models by accuracy, efficiency, transferability, and unsupervised feature learning, with CNN reaching 97.62% on the Human-ID evaluation.

  • Takeaways & Limitations

    The study recommends CNN, GRU, and BiLSTM for strong performance with low cost, and recommends unsupervised initialization for similar tasks because it extracts more generalizable features than transfer learning.

Abstract

from arXiv · show

WiFi sensing has been evolving rapidly in recent years. Empowered by propagation models and deep learning methods, many challenging applications are realized such as WiFi-based human activity recognition and gesture recognition. However, in contrast to deep learning for visual recognition and natural language processing, no sufficiently comprehensive public benchmark exists. In this paper, we review the recent progress on deep learning enabled WiFi sensing, and then propose a benchmark, SenseFi, to study the effectiveness of various deep learning models for WiFi sensing. These advanced models are compared in terms of distinct sensing tasks, WiFi platforms, recognition accuracy, model size, computational complexity, feature transferability, and adaptability of unsupervised learning. It is also regarded as a tutorial for deep learning based WiFi sensing, starting from CSI hardware platform to sensing algorithms. The extensive experiments provide us with experiences in deep model design, learning strategy skills and training techniques for real-world applications. To the best of our knowledge, this is the first benchmark with an open-source library for deep learning in WiFi sensing research. The benchmark codes are available at https://github.com/xyanchen/WiFi-CSI-Sensing-Benchmark.

I. INTRODUCTION

WiFi sensing is promising for device-free human sensing, but existing deep-learning studies lack a comprehensive public benchmark. SenseFi addresses this gap by benchmarking models and learning schemes across datasets, platforms, tasks, and practical criteria.

  • Motivation: WiFi CSI enables device-free human sensing, including occupancy detection and other applications, while offering greater privacy than cameras because it uses non-visible radio signals.The paper describes WiFi access points as sensor devices and highlights CSI-based sensing across smart environments.
  • Motivation: Model-based methods explain signal propagation and support periodic or single motions but struggle with complicated activities composed of multiple motions.The paper contrasts physical-model approaches with the complexity of activities such as human gait.
  • Research gap: Deep models address challenging WiFi sensing tasks, but current studies largely tailor networks to specific tasks instead of evaluating models comprehensively.Deep-learning approaches also require extensive labeled data and can suffer from distribution shifts caused by environmental dynamics.
  • SenseFi: SenseFi benchmarks prevalent deep models on WiFi CSI data from multiple datasets and platforms, comparing accuracy and efficiency for real-world viability.The benchmark includes raw and processed CSI data collected with Intel 5300 and Atheros CSI tools.
  • SenseFi: The benchmark evaluates transfer learning across sensing tasks and unsupervised contrastive learning to exploit unlabeled data and reduce annotation costs.These learning schemes are benchmarked across the included models.
  • SenseFi: SenseFi provides an open-source library and benchmarking codes intended to support convenient model development and evaluation for WiFi sensing research.The authors present it as the first benchmark of advanced deep models and learning schemes for WiFi sensing, to their knowledge.

II. PRELIMINARIES OF WIFI SENSING

WiFi CSI characterizes wireless-channel propagation through multipath amplitude and phase, with subcarrier resolution determined by bandwidth and the recording tool. The section presents CSI as the sensing signal captured by representative hardware platforms.

  • CSI describes wireless-channel properties after diffraction, reflection, and scattering, while MIMO and OFDM support transmission across multipath channels.
  • The channel impulse response represents multipath components through their amplitudes, phases, delays, and total count.
  • CSI records amplitude attenuation and phase shift as complex values sampled at the OFDM subcarrier level.
  • Atheros CSI Tool records 56 subcarriers at 20MHz and 114 at 40MHz, whereas Intel 5300 records 30 subcarriers at 20MHz.

C. CSI Data Transformation and Cleansing

CSI preprocessing selects representations that preserve useful sensing structure while reducing phase noise and environmental dependence. The data can be organized as spatial, temporal, and antenna dimensions for deep models.

  • Amplitude is commonly used because raw single-antenna phase is unstable, while wavelet denoising can filter high-frequency amplitude noise.
  • CSI is represented using amplitude-only data, antenna differences or ratios, and processed Doppler features; the benchmark chooses amplitude and domain-invariant BVP modalities.
  • CSI samples form a tensor over antennas, subcarriers, and time, analogous to CSI images that become CSI videos across timestamps.
  • Subcarriers encode spatial features, temporal dynamics encode temporal features, and antennas provide resolution and channel features.
  • Deep learning models automatically extract representations from labeled data, and WiFi sensing models adapt established architectures to human-motion patterns.
  • A CSI sensing model maps antenna-by-subcarrier-by-time data to task labels, with feature extractors reducing dimension while preserving the data manifold.

A. Multilayer Perceptron

MLPs provide a classic fully connected architecture, while CNNs preserve local spatial structure through shared convolution and pooling. Their differing data handling creates distinct computational and representational trade-offs for CSI sensing.

  • A. Multilayer Perceptron: MLPs transform inputs through fully connected layers and activations, ultimately mapping latent features to categorical outputs.
  • A. Multilayer Perceptron: Flattening CSI for an MLP mixes spatial and temporal dimensions, while its many parameters can cause slow convergence and high computational cost.
  • B. Convolutional Neural Network: CNNs address MLP drawbacks through weight sharing and spatial pooling, learning features by stacking convolutional kernels and pooling operations.
  • B. Convolutional Neural Network: Pooling downsamples feature maps by taking the maximum or mean within each kernel.
  • B. Convolutional Neural Network: For CSI, Conv1D extracts spatial features along subcarriers, whereas Conv2D uses two-dimensional kernels over CSI patches.

C. Recurrent Neural Network

RNNs model CSI sequences by maintaining internal memory of historical inputs, making them suitable for temporal sensing data. LSTM and bidirectional variants address long-term dependencies and expand sequence modeling capacity.

  • C. Recurrent Neural Network: RNNs memorize arbitrary-length input sequences through internal memory trained with back-propagation through time.
  • C. Recurrent Neural Network: A vanilla RNN generates each hidden state from the current CSI frame using shared input and hidden-state matrices.
  • C. Recurrent Neural Network: Although RNNs capture temporal dynamics, vanishing gradients prevent them from capturing long-term CSI dependencies.
  • D. Variants of RNN (LSTM): LSTM mitigates gradient instability with memory cells and gates controlling output, input, and forgetting across the sequence.
  • D. Variants of RNN (LSTM): Figure 3 illustrates how CSI data is processed by MLP, CNN, RNN, and Transformer architectures.
  • D. Variants of RNN (LSTM): Bidirectional LSTM processes sequences in forward and backward directions and concatenates the resulting features.

E. Recurrent Convolutional Neural Network

Recurrent convolutional models combine convolutional spatial-feature extraction with recurrent temporal modeling for WiFi sensing. The benchmark uses CNN-GRU because it performs similarly to alternatives with fewer parameters.

  • CNN-GRU combines Conv1D and GRU to extract spatial and temporal features.
  • CNN-GRU has also been applied to CSI-based human gesture recognition in Widar.
  • The benchmark selects CNN-GRU over Conv1D with BiLSTM because the models perform similarly while CNN-GRU uses fewer parameters.
  • Transformer attention computes pairwise patch relationships through dot products, normalizes weighting, and repeats attention across multiple heads.

G. Generative Models

This section surveys generative and alternative learning strategies for WiFi sensing. It emphasizes reducing dependence on massive labeled datasets and addressing environmental variation through transfer, unsupervised, few-shot, and ensemble approaches.

  • G. Generative Models: Generative models learn CSI data distributions rather than directly performing classification.
  • G. Generative Models: GANs generate real-like labeled CSI samples for new environments from models trained in existing environments.
  • Supervised learning achieves high performance but requires tremendous labeled data, making collection and annotation a practical bottleneck.
  • Few-shot learning trains with several samples per category and supports WiFi gesture recognition and human identification.
  • Transfer learning pretrains models in one domain and fine-tunes them in another, while domain adaptation addresses distribution shifts between distinct CSI environments.
  • Unsupervised learning learns representations without labels, but has not been extensively exploited in WiFi sensing.

V. EMPIRICAL STUDIES OF DEEP LEARNING IN WIFI SENSING: A BENCHMARK

SenseFi benchmarks deep architectures across four CSI datasets spanning multiple sensing tasks and hardware platforms. Results show that model effectiveness varies by dataset, while simpler architectures can provide strong accuracy and practical complexity.

  • Datasets and setup: The benchmark evaluates deep models on four CSI datasets, including two public datasets and two newly collected datasets from different CSI platforms.
  • Datasets and setup: UT-HAR contains seven activity categories from Intel 5300 NIC data, while Widar contains 22 gesture categories and 43K samples.
  • Datasets and setup: NTU-Fi includes human activity recognition and human identification tasks, with 114 subcarriers per antenna pair and perfectly segmented CSI samples.
  • Overall comparison: 98.11% accuracy is achieved by ResNet-18 on UT-HAR, with CNN-5 achieving the second-best result.
  • Overall comparison: MLP, GRU, and CNN show stable and superior performance across the benchmark comparisons.
  • Overall comparison: Very deep ResNets do not outperform simple CNNs on NTU-Fi, and performance does not increase consistently with network depth.
  • Complexity and practicality: The transformer performs poorly with limited NTU-Fi Human-ID training data and has very high computational complexity relative to comparable models.
  • Complexity and practicality: CNN-5, GRU, BiLSTM, and CNN-GRU offer favorable combinations of parameter size and accuracy for edge inference.

E. Evaluations of Learning Schemes

The benchmark evaluates transfer and unsupervised learning as alternatives to conventional supervised training. CNN provides the strongest reported transfer and unsupervised-learning results on the NTU-Fi evaluation tasks.

  • The study evaluates transfer learning and unsupervised learning strategies on WiFi sensing models.
  • Transfer learning: Transfer learning pretrains models on NTU-Fi HAR and fine-tunes a new classifier on NTU-Fi Human-ID.
  • Transfer learning: 96.35% accuracy is achieved by the CNN feature extractor on the Human-ID transfer task.
  • Transfer learning: RNN, CNN+GRU, and ViT achieve 57.84%, 51.73%, and 66.20%, respectively, on the transfer-learning evaluation.
  • Unsupervised learning: Unsupervised learning trains parallel networks using only CSI data, then trains an independent classifier on their fixed parameters.
  • Unsupervised learning: 97.62% accuracy is achieved by CNN on the NTU-Fi Human-ID task after unsupervised training on NTU-Fi HAR.

F. Analysis

The experiments compare model convergence, transfer learning, generalization, and optimizer behavior across WiFi sensing datasets. Results show fast convergence for CNNs and MLPs, improved stability with transfer learning, overfitting in deep ResNets on Widar, and optimizer-dependent training behavior.

  • Convergence of Deep Models: CNN converges within 25 epochs across four datasets, while MLP also converges quickly; transformers require more training epochs.The transformer’s greater parameter count contributes to its slower convergence.
  • How Transfer Learning Matters: Pre-trained parameters make convergence more stable across similar NTU-Fi sensing tasks, indicating transferable feature extractors.The comparison uses NTU-Fi Human-ID initialized from NTU-Fi HAR against randomly initialized models.
  • Poor Performance of Deep CNN on Widar: 17.91%, 19.47%, and 14.47% are the Widar accuracies of ResNet-18, ResNet-50, and ResNet-101, respectively.Despite nearly 100% training accuracy, testing accuracy remains under 20%, indicating overfitting in the cross-domain setting.
  • Choices of Optimizer: Adam accelerates convergence but causes fluctuating losses in ResNet-50 and ResNet-101, whereas ResNet-18 converges stably.The instability is associated with changing WiFi data values and Adam’s adaptive learning rate.

VI. DISCUSSIONS AND SUMMARY

The discussion recommends shallow, computationally efficient models and careful optimizer selection for WiFi sensing. It also identifies transfer learning and unsupervised initialization as useful strategies when tasks and sensing platforms are sufficiently similar.

  • Model Choices: CNN, GRU, and BiLSTM are recommended for their high performance, low computational cost, and small parameter size.Shallow models perform well across activity recognition, gesture recognition, and human identification, while very deep models face overfitting in cross-domain scenarios.
  • Optimization: Adam or SGD are recommended: Adam converges quickly but may be unstable, while SGD is more secure but requires manual learning-rate and momentum tuning.The optimizer choice therefore involves a speed-versus-stability and tuning trade-off.
  • Advice on Transfer Learning Applications: Transfer learning is recommended for similar tasks using the same CSI platform because pre-trained parameters provide good initialization and better generalization.CNN, MLP, and BiLSTM are reported to have superior transferability.
  • Advice on Unsupervised Learning: Unsupervised learning is recommended for similar tasks because it extracts more generalizable features than transfer learning; CNN, MLP, and ViT are generally suitable.The recommendation targets model initialization rather than replacing task-specific evaluation.

VII. GRAND CHALLENGES AND FUTURE DIRECTIONS

The paper identifies future needs around data efficiency, lightweight deployment, multimodal and cross-modal sensing, trustworthy models, complex behavior analysis, and interpretability. These directions address limited data, real-time processing, richer sensing, security, difficult behaviors, and physical understanding.

  • Data-efficient learning: Data-efficient learning remains important because CSI collection is expensive and current transferability is weakly evaluated in simple testing scenarios.The paper points to meta-learning and zero-shot learning for robust features across environments and tasks.
  • Model compression or lightweight model design: Model compression and lightweight architectures are future priorities for real-time WiFi sensing applications such as vital-sign monitoring.Pruning, quantization, and distillation are identified as ways to decrease model size.
  • Multi-modal learning: Multimodal learning can fuse WiFi and visual sensing to learn joint features and choose reliable modalities for robust 24/7 sensing.WiVi is cited as an example integrating WiFi sensing and visual recognition.
  • Cross-modal learning: Cross-modal learning can use CSI to supervise or reconstruct visual information, with Wi2Vi generating video frames from CSI data.The paper also describes pose estimation supervised by OpenPose landmarks.
  • Model robustness and security for trustworthy sensing: Trustworthy WiFi sensing requires addressing privacy leakage and adversarial attacks that can cause incorrect smart-appliance actions.Proposed directions include adversarial training, CSI quantization, privacy protection, and federated learning.
  • Complicated human activities and behaviors analytics: Existing methods handle single activities or gestures better than complicated behavior analytics, such as detecting anomalous routines related to Alzheimer’s risk.Interpretability is also sought to connect data-driven models with physical sensing models and inspire new physical theories.
Loading 2207.07859v3…