AI-Driven Wi-Fi Biometrics WhoFi Tracks Humans Behind Walls with 95.5% Accuracy

Researchers have introduced WhoFi, an AI-powered deep learning pipeline that leverages Wi-Fi Channel State Information (CSI) for person re-identification (Re-ID), achieving a remarkable 95.5% Rank-1 accuracy on the NTU-Fi dataset.

Traditional visual Re-ID systems, reliant on convolutional neural networks (CNNs) and features like color histograms or Histograms of Oriented Gradients (HOG), falter under occlusions, varying illumination, and viewpoint changes.

Revolutionizing Surveillance

WhoFi circumvents these issues by exploiting non-visual modalities, where Wi-Fi signals penetrate walls and occlusions, capturing biometric signatures from internal body structures such as bones and organs via signal distortions in CSI matrices.

This approach not only enhances robustness in unconstrained environments but also ensures privacy, as it avoids capturing identifiable visual data.

The core innovation lies in processing CSI-derived amplitude and phase data through advanced sequence modeling architectures, including Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Transformer encoders, trained with an in-batch negative loss to optimize embedding spaces for similarity matching.

WhoFi’s modular architecture begins with data preprocessing, applying Hampel filters for amplitude outlier removal and linear sanitization for phase offsets, followed by augmentations like Gaussian noise addition, amplitude scaling, and time shifts to bolster model generalization.

The encoder module extracts latent representations from sequential CSI inputs: LSTMs handle temporal dependencies via stacked hidden units with dropout regularization, Bi-LSTMs incorporate bidirectional context for enhanced pattern recognition, and Transformers utilize multi-head self-attention with positional encodings to model long-range correlations efficiently.

Deep Learning Architectures

A signature module then projects these encodings into a normalized hypersphere using L2 normalization, facilitating cosine similarity computations.

Trained on batches comprising query and gallery samples, the in-batch negative loss maximizes diagonal similarities in the matrix (positive pairs) while minimizing off-diagonals (negatives) via cross-entropy, enabling scalable learning without explicit pair labeling.

Empirical evaluations on the NTU-Fi dataset, featuring CSI amplitudes from 14 subjects across 840 samples collected via MIMO-OFDM setups with 114 subcarriers and 2000 packets per sample, underscore the Transformer’s dominance.

It outperforms LSTM (77.7% Rank-1, 56.8% mAP) and Bi-LSTM (84.5% Rank-1, 61.2% mAP) with 95.5% Rank-1, 98.1% Rank-3, 99.1% Rank-5, and 88.4% mean Average Precision (mAP).

Ablation studies reveal counterintuitive insights: amplitude filtering slightly degrades performance by eliminating discriminative noise, while augmentations boost LSTM and Bi-LSTM but yield marginal gains for Transformers.

Varying packet sizes show Transformers thrive on longer sequences (up to 2000 packets) due to self-attention’s efficiency, unlike LSTMs prone to vanishing gradients.

Deeper architectures (3 layers) marginally aid recurrent models but induce overfitting in Transformers, affirming single-layer efficiency.

This work, implemented in PyTorch with Adam optimization and StepLR scheduling over 300 epochs, establishes a reproducible baseline for wireless biometrics.

By harnessing radio frequency interactions for Re-ID, WhoFi paves the way for applications in smart environments, from security systems to healthcare monitoring, where visual limitations hinder deployment.

Future extensions could integrate phase data or multi-modal fusion, further elevating accuracy in diverse scenarios.

Get Free Ultimate SOC Requirements Checklist Before you build, buy, or switch your SOC for 2025 - Download Now