Wi-Spike: A Low-power WiFi Human Multi-action Recognition Model with Spiking Neural Networks

1. Problem Statement

While WiFi-based Human Action Recognition (HAR) offers non-intrusive and privacy-preserving advantages over cameras and wearables, existing solutions face three critical limitations:

• High Energy Consumption: Most state-of-the-art models rely on deep learning architectures (CNNs, RNNs, Transformers) that require heavy continuous computation, making them unsuitable for battery-constrained edge devices (e.g., IoT nodes).
• Focus on Single Actions: Current research predominantly targets simple, singular actions (e.g., walking, jumping). There is a significant lack of robust frameworks capable of recognizing multi-action scenarios (complex, overlapping, or sequential combinations of actions) which are common in real-world daily life.
• Accuracy-Efficiency Trade-off: Existing approaches often treat recognition accuracy and energy efficiency as conflicting objectives, failing to provide a unified framework that achieves both high performance and low power consumption.

2. Methodology: Wi-Spike

The authors propose Wi-Spike, a bio-inspired framework based on Spiking Neural Networks (SNNs) designed to leverage the event-driven and low-power nature of biological neurons for WiFi Channel State Information (CSI) processing.

Core Architecture

The model processes raw WiFi CSI signals through four main stages:

1. Spike Encoding: Continuous CSI data is normalized and transformed into discrete spike trains. This mimics the event-driven nature of biological neurons, converting static spatial features into temporal spike sequences.
2. Spike Feature Extraction:
   • Spiking Convolutional Layers: Extract spatial and temporal patterns from the spike-encoded CSI data.
   • Temporal Attention Mechanism: A novel module designed specifically for SNNs. It uses localized pooling (average and max) and a learnable transformation to generate attention weights. This layer adaptively emphasizes salient time steps (e.g., specific motion phases) while suppressing irrelevant noise and multipath interference.
3. Spike Feature Encoding: Spiking pooling layers reduce spatial dimensions while preserving critical temporal information, maintaining the sparse, low-power nature of the network.
4. Voting Classification:
   • Spiking Fully Connected Layers: Process the encoded features.
   • Spiking Voting Layer: Aggregates spike activities over time to calculate a population firing rate. The class with the highest firing rate (most active neurons) is selected as the final prediction.

Training Strategy

To optimize the hybrid SNN, the authors employ a hybrid loss function:

• Mean Squared Error (MSE): Aligns the predicted firing rates with ground truth labels.
• Supervised Contrastive Loss (SCL): Enhances feature discriminability by pulling embeddings of similar actions closer and pushing dissimilar ones apart in the feature space.

3. Key Contributions

1. Novel SNN Framework for WiFi: Wi-Spike is the first work to introduce a bio-inspired spiking neural network specifically for WiFi action recognition, replacing traditional Artificial Neural Network (ANN) neurons with spiking neurons to drastically reduce power consumption.
2. Multi-Action Recognition Challenge: The paper defines and addresses the "Multi-Action" sensing problem (recognizing composite actions like "walk + run" or "fall + wave"). The authors constructed new datasets (Double and Triple action variants) by concatenating single-action samples to evaluate model robustness in complex scenarios.
3. Temporal Attention for SNNs: The design of a spike-based temporal attention layer that allows the network to focus on discriminative spatio-temporal patterns within CSI signals, improving the model's ability to disentangle complex motion patterns.
4. Energy-Efficient State-of-the-Art: The model achieves a new state-of-the-art balance between accuracy and energy efficiency, proving that SNNs are viable for real-time, edge-based WiFi sensing.

4. Experimental Results

The model was evaluated on three benchmark datasets (UT-HAR, NTU-Fi-HAR, NTU-Fi-HumanID) and their constructed multi-action variants.

• Single-Action Recognition:

  • Wi-Spike achieved competitive accuracy, reaching 96.97% on NTU-Fi-HAR, 95.58% on NTU-Fi-HumanID, and 96.51% on UT-HAR.
  • It outperformed ViT and CNN+GRU models and remained comparable to top-performing CNNs and LSTMs.

• Multi-Action Recognition:

  • Wi-Spike demonstrated superior robustness in complex scenarios. On the NTU-Fi-HAR Double dataset, it achieved 95.83% accuracy, outperforming the next best model (CNN) by ~4%.
  • As the number of actions increased (from 1 to 3), Wi-Spike maintained the highest accuracy with the lowest variance, whereas traditional models (like ViT) suffered significant performance drops.

• Energy Efficiency:

  • Power Consumption: Wi-Spike reduced energy costs by at least 50% compared to other methods.
  • Theoretical Cost: On the NTU-Fi-HAR Double dataset, Wi-Spike consumed only 0.07 pJ per inference, compared to 2.33 pJ for ViT and 0.25 pJ for ResNet18.
  • Efficiency Ratio: It requires only 1/33rd of the energy of a ViT model to achieve similar or better sensing tasks.

• Firing Rate Analysis: The model operated within an optimal firing rate range (0.1–0.3), ensuring stable, biologically plausible dynamics without over-activation.

5. Significance

This work represents a paradigm shift in WiFi sensing by moving away from energy-intensive continuous computation toward event-driven, sparse processing.

• Edge Deployment: By drastically lowering power consumption while maintaining high accuracy, Wi-Spike makes long-term, real-time human activity monitoring feasible on resource-constrained edge devices (e.g., smart home hubs, wearable sensors).
• Complex Behavior Understanding: It successfully addresses the gap in recognizing complex, multi-action human behaviors, which is crucial for advanced applications in elderly care, healthcare monitoring, and smart environments.
• Sustainability: The approach offers a sustainable solution for the growing field of pervasive computing, aligning high-performance AI with green computing principles.