EfficientFi: Towards Large-Scale Lightweight WiFi Sensing via CSI Compression

Motivation

• CSI sensing streams have high dimensionality and sampling rates, creating communication burdens that can hinder basic WiFi Internet access.The paper identifies compression as necessary for transmitting large-scale sensing data from users to cloud servers.Source: §I · pp. 1–2
• Cloud-based sensing additionally requires both accurate recognition and CSI reconstruction for IoT data logging and incremental model learning.These requirements extend beyond simply transmitting compressed CSI for recognition.Source: §I · p. 2

EfficientFi Framework

• EfficientFi provides CSI compression, CSI restoration, and CSI-based recognition to bridge single-user sensing and large-scale WiFi sensing.The framework assigns feature extraction and compression to WiFi routers, while cloud servers perform recognition and reconstruction.Source: §I · p. 2
• EfficientFi uses a quantized low-dimensional CSI representation and joint learning to make compressed features discriminative for recognition.Its multi-task, end-to-end design jointly supports compression, restoration, and recognition.Source: §I · p. 2

Evaluation

• Over 1,700× communication-burden reduction was achieved without degrading recognition performance in WiFi activity-recognition and human-identification experiments.The experiments evaluate the framework in real-world WiFi sensing settings.Source: §I · p. 2

WiFi Sensing Systems and Applications

• WiFi sensing systems use CSI for applications including occupancy detection, crowd counting, activity recognition, respiration monitoring, person identification, and gesture recognition.CSI can be extracted from WiFi hardware at increasing subcarrier resolutions and directly at access points.Source: A. WiFi Sensing Systems and Applications · p. 2

Recognition Models

• Model-based methods use propagation and signal-processing theories, whereas learning-based methods learn statistical or deep-neural-network patterns from data.The paper focuses more directly on learning-based recognition models.Source: B. Deep Learning Models for WiFi Sensing · p. 2

Deep Learning Models for WiFi Sensing

• Deep learning models capture more complicated WiFi sensing patterns and improve recognition granularity, but require massive data collection and greater computation overhead.These requirements commonly motivate use of cloud servers for WiFi sensing.Source: B. Deep Learning Models for WiFi Sensing · pp. 2–3
• Prior deep-learning approaches apply temporal, spatial, recurrent, convolutional, and Siamese architectures to activity recognition, gesture recognition, human identification, and localization.EfficientFi combines deep CSI compression, quantization, and recognition within a large-scale sensing framework.Source: B. Deep Learning Models for WiFi Sensing · p. 3

• Traditional CSI feedback methods use compressive sensing or LASSO, while learned autoencoders improve recovery quality from training samples.Source: C. CSI Compression in Communication · p. 3
• CSI records amplitude attenuation and phase shifts of multipath components for each antenna pair and OFDM subcarrier.Source: A. Channel State Information · p. 3
• WiFi sensing can compress only task-relevant CSI components, such as subcarrier amplitudes for human activity recognition, rather than the whole matrix.Source: A. Channel State Information · p. 3
• EfficientFi extends low-dimensional CSI compression by requiring the compressed representation to be discrete for communication and cloud restoration.Source: A. Channel State Information · p. 3

WiFi Sensing

• Human activities alter wireless-signal propagation, producing CSI patterns that learning or analytic methods can use for recognition.Source: B. WiFi Sensing and Limitations · p. 3
• EfficientFi bridges edge and cloud computing by learning a discriminative compressed CSI feature space for large-scale WiFi sensing.Source: B. WiFi Sensing and Limitations · p. 4

System Overview

• At the WiFi AP, a lightweight CNN extracts features and quantizes them by nearest-vector lookup in a CSI codebook.Source: A. System Overview · p. 4
• EfficientFi defines a feature extractor, decoder, and task classifier, jointly optimizing extraction and classification while reconstructing the original CSI sample.Source: A. System Overview · p. 4

Quantization

• Each continuous encoder feature is assigned to its nearest CSI codebook vector, producing a one-hot quantized representation.Source: B. Discrete Representation Learning… · p. 4

Restoration

• At the cloud server, the quantized representation is reversed into a reconstructed feature and decoded into CSI for database storage.Source: B. Discrete Representation Learning… · p. 4

Recognition

• The classifier uses reconstructed CSI features to predict task labels such as human activity categories.Source: C. Joint Learning of Quantized… · p. 4

Optimization Challenges

• Training must learn a comprehensive codebook and overcome the non-differentiability of quantized feature generation.Source: D. Learning and Optimization · p. 5

Optimization

• VQ-VAE addresses these difficulties by learning codebook parameters with back-propagation and using straight-through gradient estimation for end-to-end training.Source: D. Learning and Optimization · p. 5

Learning Objectives

• The reconstruction objective trains encoder and decoder parameters through mean squared error, with a straight-through estimator linking continuous and discrete features.Source: D. Learning and Optimization · p. 5
• The codebook objective minimizes the distance between continuous encoder outputs and quantized features using vector quantization.Source: D. Learning and Optimization · p. 5
• Joint feature-extractor and classifier training makes the compressed feature space discriminative for WiFi sensing tasks.Source: D. Learning and Optimization · p. 5

Training

• EfficientFi’s overall loss comprises separate optimization processes for the model parameters and codebook.Source: E. Algorithm Summary · p. 5

Deployment

• Deployment places the feature extractor at the edge, the decoder and classifier at the cloud, and stores the codebook at both sides.Source: E. Algorithm Summary · p. 5

Design Rationale

• EfficientFi’s learned codebook captures CSI patterns for reconstruction, while multi-task learning integrates recognition and reconstruction objectives to learn semantic features.Source: E. Algorithm Summary · p. 5

Algorithm

• The training algorithm extracts and quantizes features, reconstructs CSI, predicts sensing results, and updates network parameters and the codebook iteratively.Source: E. Algorithm Summary · p. 5

Architecture

• The experimental network architecture includes convolution, transposed-convolution, max-pooling, and max-unpooling operations.Source: Table I · p. 6

System Design

• Experiments use two TP-Link N750 APs with three antenna pairs, 40 MHz bandwidth, and 114 CSI subcarriers per timestamp.A local server simulates cloud computing.Source: A. Experimental Setup · p. 6

Data Collection

• 500 Hz sampling produces 3 × 114 × 500 CSI samples per second for six activities performed by 20 subjects.Each activity category contains 400 samples.Source: A. Experimental Setup · p. 6

Network Implementation

• The network uses four convolutional or transposed-convolutional layers for feature extraction and decoding, plus a two-layer softmax classifier.The codebook uses D = 256, while K varies from 64 to 1024 and λ = 0.5.Source: A. Experimental Setup · p. 6

Criterion

• Evaluation compares communication cost, NMSE reconstruction error, and recognition accuracy.The original communication cost is 1.368Mb/s, and NMSE is reported in decibels.Source: A. Experimental Setup · p. 6

Baselines

• Compression baselines include LASSO, BM3D-AMP, CSINet, and a vanilla auto-encoder, while HAR and gait identification use task-specific comparison methods.E-eyes and CARM support HAR comparison; WiWho and AutoID support gait recognition.Source: A. Experimental Setup · pp. 6–7

Overall comparison

• 334 compression rate is achieved with K = 256 while maintaining satisfactory reconstruction error for both HAR and human identification.EfficientFi outperforms LASSO, BM3D-AMP, and CSINet in the reported comparisons.Source: B. Evaluation on HAR and Human-ID · p. 7

HAR

• 20% around improvement over E-eyes and CARM is reported for human activity recognition.The uncompressed encoder-only accuracy is 98.3%, with accuracy declining as compression increases.Source: B. Evaluation on HAR and Human-ID · p. 7

Human-ID

• 83% around accuracy is retained for person identification after compression, compared with 67.3% for WiWho and 77.6% for AutoID.The best reported compressed result occurs at compression rate 66.8, while higher compression reduces accuracy.Source: B. Evaluation on HAR and Human-ID · p. 7

Summary

• EfficientFi achieves high-fidelity reconstruction and high recognition accuracy for both wireless activity recognition and person identification.The conclusion summarizes preservation of classification performance after feature compression and transmission.Source: B. Evaluation on HAR and Human-ID · p. 7

Incremental learning

• The reconstructed data is used to fine-tune the classifier with high-confidence samples for improved recognition performance.The result is reported for the Human-ID evaluation in Table III.Source: C. Incremental Learning via… · pp. 7–8, Table III · p. 8
• 89.5% accuracy is obtained after fine-tuning with reconstructed CSI and a 90% confidence threshold.Half of the original testing data is used for fine-tuning and half for evaluation.Source: C. Incremental Learning via… · pp. 7–8

Hyperparameter sensitivity

• With K = 64, increasing embedding dimension D reduces accuracy variance across settings, with D = 64 producing 95%–98% accuracy variation.The analysis also varies the weight λ and focuses on recognition accuracy because NMSE is acceptable across settings.Source: D. Hyperparameter Sensitivity · p. 8, Figure 4 · p. 8

Inference Time

• 2.1ms is required to compress one second of CSI at 500Hz, versus 251ms for LASSO and 747ms for BM3D-AMP.CSINet requires 5.1ms per CSI timestamp, so its cost scales with sampling rate.Source: E. Inference Time · p. 8

Feature visualization

• T-SNE shows raw CSI activities can overlap, whereas quantized features form separated discriminative clusters.The reconstructed feature preserves discriminability before the classifier further aligns the latent space.Source: F. T-SNE Visualization · p. 8

Conclusion

• EfficientFi combines quantized representation learning with joint recognition learning in an end-to-end offline framework for online edge-cloud sensing.Experiments report state-of-the-art performance against classic compressive sensing and deep compression methods.Source: §VI · p. 8