Source-linked AI summary
XAI2CSI: Interpreting CSI with eXplainable AI for Human Activity Recognition
Idio Guarino, Alfredo Nascita, Domenico Ciuonzo, Damiano Carra, Antonio Pescapé
TL;DR
CSI enables unobtrusive device-free HAR, but DL models generalize poorly across changing users, environments, and devices, and their black-box decisions remain insufficiently understood. XAI2CSI applies SAGE to quantify temporal, spectral, and spatial CSI contributions under nominal and cross-context evaluations. The analysis finds limited robustness driven by reliance on context-specific patterns that are misinterpreted when deployment conditions change.
Problem
CSI-based DL HAR models struggle to generalize across users, environments, and hardware, while model decisions and failures under unseen conditions remain insufficiently understood.
Method
XAI2CSI applies the model-agnostic SAGE method to quantify temporal, spectral, and spatial CSI contributions to HAR decisions under nominal and cross-context evaluations.
Results
The analysis finds limited robustness under unseen conditions because models over-rely on context-specific CSI patterns and misinterpret signal dynamics when deployment conditions change.
Takeaways & Limitations
Explainability provides a principled way to understand DL behavior in Wi-Fi sensing and guide more transparent, interpretable, and robust systems.
Abstract
from arXiv · showhide
Wi-Fi Channel State Information (CSI) has emerged as a key enabler for device-free Human Activity Recognition (HAR), enabling low-cost, unobtrusive sensing using existing communication infrastructure. However, Deep Learning (DL) models trained on CSI data often struggle to generalize across users, environments, and device setups due to the context sensitivity of wireless propagation. Despite this challenge, limited attention has been devoted to understanding model decisions and generalization failures. This paper introduces XAI2CSI, a framework that leverages eXplainable Artificial Intelligence (XAI) to analyze DL-based CSI sensing systems. XAI2CSI employs SAGE, a model-agnostic explainability method, to quantify temporal, spectral, and spatial CSI contributions to HAR decisions under nominal and cross-context evaluations on IEEE 802.11ax data. Our analysis reveals that the considered DL model exhibits limited robustness to unseen conditions due to an over-reliance on context-specific CSI patterns, causing models to misinterpret the underlying signal dynamics when deployment conditions change. The proposed methodology and findings provide a reference framework to explore alternative solutions and guide the development of robust, transparent Wi-Fi sensing systems.
I. INTRODUCTION
CSI enables low-cost, device-free HAR with existing Wi-Fi infrastructure, but context-sensitive propagation limits DL generalization and leaves model failures difficult to explain. XAI2CSI addresses this gap by analyzing CSI contributions across deployment contexts.
- I. INTRODUCTION: CSI supports low-cost, non-intrusive HAR by sensing movement-induced disturbances in radio-frequency propagation.It avoids wearable-device limitations including battery constraints, discomfort, and complex deployment.
- I. INTRODUCTION: CSI-based DL models may fail across users, environments, and device setups because wireless propagation is highly context-sensitive.Multipath reflections create setting-specific CSI fingerprints that can change with minor environmental differences.
- I. INTRODUCTION: DL models operate as black boxes, obscuring which CSI components drive predictions and how their contributions change across deployment contexts.The paper identifies why CSI-based HAR models fail under unseen conditions as a key open question.
- I. INTRODUCTION: XAI2CSI applies model-agnostic SAGE to quantify temporal, spectral, and spatial CSI contributions under nominal and cross-context evaluations.The framework evaluates CSI-based HAR models using IEEE 802.11ax data and analyzes failures across users, environments, and receiver positions.
- I. INTRODUCTION: The analysis attributes limited robustness to over-reliance on context-specific CSI patterns that misrepresent signal dynamics after deployment conditions change.The findings are intended to support robust, transparent, and interpretable Wi-Fi sensing systems.
II. RELATED WORKS
Prior CSI-HAR research has expanded in sensing capability and signal representations, but remains constrained by hardware access, computational trade-offs, limited cross-context evaluation, and opaque DL decisions. Table I positions this work within those gaps.
- II. RELATED WORKS: CSI-based human sensing spans presence detection, HAR, and vital-sign monitoring, progressing from coarse metrics to fine-grained subcarrier data.DL has further advanced these sensing capabilities.
- II. RELATED WORKS: Table I summarizes representative HAR works by Wi-Fi version, signal representation, cross-evaluation scenario, and explainability.The table is presented as the positioning context for this work.
- II. RELATED WORKS: A lack of native CSI extraction on commodity devices has tied early research to specialized tools and outdated 802.11n hardware.Newer Nexmon and AX-CSI frameworks support 802.11ac and 802.11ax, but many studies still use legacy tools.
- II. RELATED WORKS: Amplitude is widely used for simplicity and real-time suitability, whereas phase differences, Doppler, STFT, and body velocity provide richer motion information with added costs and sensitivities.The alternatives may require calibration, increase computation, and suffer orientation or multipath distortions.
- II. RELATED WORKS: Generalization remains a major challenge because CSI depends strongly on users, environments, and hardware, while many studies evaluate models within the same context.Adaptation methods can improve cross-domain performance but require new data collection in each context.
III. METHODOLOGY
The methodology models CSI as structured temporal, spatial, and spectral inputs for HAR, then uses SAGE to measure how feature subsets affect predictive loss and global model performance.
- III. METHODOLOGY: CSI in a MIMO system is represented at time t as H(t) ∈ C^Ntx×Nrx×F across transmit-receive antenna pairs and subcarriers.The representation captures channel responses that reflect static room factors and dynamic human movement.
- III. METHODOLOGY: The HAR model compares amplitude, phase-difference, and combined amplitude-phase CSI representations as inputs.The setup uses a 1 × Nrx MIMO configuration because different activities perturb wireless propagation differently.
- III. METHODOLOGY: Each CSI frame preserves spatial, spectral, and temporal structure before the DL model outputs activity-class probabilities.The predicted activity is the class with maximum model probability, and training uses categorical cross-entropy.
- III. METHODOLOGY: SAGE quantifies each CSI feature’s global contribution by comparing predictive loss with and without feature subsets.Unlike local explainability methods, it measures global loss degradation across feature distributions.
- III. METHODOLOGY: SAGE assigns feature contributions through Shapley values of a predictive-power function based on conditional model expectations.The conditional prediction marginalizes over features not included in the evaluated subset.
IV. EXPERIMENTAL SETUP
The experiments use modern IEEE 802.11ax CSI data collected across users, environments, days, and receiver settings, with preprocessing designed to reduce dimensionality and normalize inputs before DL evaluation.
- IV. EXPERIMENTAL SETUP: The dataset contains 80-second activity traces collected across three users, three furniture-rich environments, and two consecutive days.Each trace corresponds to a user, activity, environment, and day.
- IV. EXPERIMENTAL SETUP: CSI was captured with the AX-CSI tool on IEEE 802.11ax hardware using a 1 × 4 SIMO setup in the 5GHz band at 150Hz.The setup supports evaluation across different receiver settings.
- IV. EXPERIMENTAL SETUP: Preprocessing downsamples CSI to 80MHz and 1024 subcarriers, removes non-informative bins to reach 994 subcarriers, and normalizes per receiver antenna.These steps reduce computational load while standardizing receiver inputs.
- IV. EXPERIMENTAL SETUP: SAGE computation flattens each frame while preserving its original structure for grouping by subcarrier, antenna, or time index.Conditional expectations use 40 training samples per class with a convergence threshold of 0.05.
V. EXPERIMENTAL EVALUATION
The evaluation tests CSI-based HAR under cross-user, cross-environment, and cross-node shifts, then uses SAGE to identify spatial, temporal, and spectral sources of generalization failures. Performance and feature-importance analyses show that context-dependent CSI patterns, especially amplitudes, can produce severe unseen-context errors.
- Model performance: 88–98% accuracy is achieved consistently by Mall models, with amplitude-based inputs outperforming phase-only input.These results provide an upper-bound reference using models trained on all settings.
- Model performance: At most 65% accuracy is obtained by M∗ X under cross-user evaluation, while cross-environment accuracy remains at or below 28%.Amplitude-based inputs are strongest for cross-user shifts, whereas phase-based inputs perform slightly better under cross-environment and cross-node shifts.
- Misclassification patterns: Early fusion of amplitude and phase difference resolves near-perfectly the Empty-class error but only partially mitigates confusion for Run and Sitting.Standalone amplitude or phase inputs make Run and Sitting especially difficult, with Sitting never correctly classified in the described setting.
- SAGE analysis: SAGE separates jointly correct samples Ts from samples Tu that Mall classifies correctly but M∗ X misclassifies, enabling analysis of features associated with cross-user failures.The analysis examines receiving antennas, packet arrival order, and subcarriers.
- Spatial features: Mall maintains a stable antenna hierarchy dominated by amplitude, whereas M∗ X assigns unstable antenna importance across unseen users and can reverse which antenna supports or drives errors.For U1, M∗ X over-relies on Rx,1 and underweights Rx,2; for U3, Rx,4 supports correct predictions but drives misclassifications despite being least influential for Mall.
- Temporal and spectral features: M∗ X misinterprets temporal and spectral structure on Tu: previously helpful packets become harmful, and nearly all subcarriers receive negative importance.Mall distributes positive importance across a wide spectral range, while the DC subcarrier has no effect for either model.
VI. DISCUSSION AND CONCLUSIONS
XAI2CSI uses SAGE to examine CSI-based HAR under changing deployment contexts. The analysis links cross-user generalization failures to reliance on user-specific patterns that are misinterpreted for new users.
- Performance drops significantly when deployment contexts change, showing that cross-domain robustness in Wi-Fi sensing remains an open challenge.
- XAI2CSI applies SAGE to quantify temporal, spectral, and spatial CSI contributions to model decisions in cross-user evaluation.
- Generalization failures stem from over-reliance on user-specific CSI patterns that models misinterpret when encountering a new user.