Channel State Information Authentication

• CSI authentication is a physical-layer technique that leverages channel frequency responses and RF imperfections to uniquely verify wireless signals.
• It employs methods such as similarity testing, sliding window analysis, and deep learning architectures like CNNs and Siamese networks for robust verification.
• Effective implementation requires robust preprocessing, temporal consistency, and defenses against replay and adversarial attacks in dynamic environments.

Channel State Information (CSI) authentication is a class of physical-layer authentication techniques that uses the channel frequency response observed on OFDM subcarriers to verify a transmitter, a copresent device, or, in some systems, a biometric subject. Across the recent literature, CSI is treated either as a link-specific descriptor of propagation, whose temporal and spatial evolution can be matched against enrolled observations, or as a carrier of device-specific micro-signals caused by RF circuitry imperfections. This has led to several distinct but related lines of work: source authentication for Wi-Fi management frames, mobile device authentication, copresence detection, RF fingerprinting, open-set transmitter rejection, and CSI-based biometrics (Guo et al., 28 Aug 2025, Jiang et al., 2012).

1. Foundations of CSI authentication

CSI describes the complex channel response on each subcarrier. In one formulation, the response on subcarrier $k$ is written as $H_k = \mathfrak{Re}\{H_k\} + j\mathfrak{Im}\{H_k\}$, with magnitude and phase given by

$$M_k = \sqrt{\mathfrak{Re}\{H_k\}^2 + \mathfrak{Im}\{H_k\}^2}, \qquad \phi_k = \mathrm{atan}\left(\frac{\mathfrak{Im}\{H_k\}}{\mathfrak{Re}\{H_k\}}\right).$$

CSI across subcarriers can also be related to the channel impulse response through a DFT relationship, so a single CSI measurement encodes both multipath structure and environmental context (Fomichev et al., 2021).

The appeal of CSI for authentication is tied to several properties repeatedly emphasized in the literature. Early Wi-Fi work exploited its rapid spatial decorrelation, relative independence of transmit power, and much richer dimensions than RSS, arguing that a spoofing transmitter cannot easily reproduce the legitimate channel unless it is effectively collocated with the legitimate sender. On that basis, CSI was used to authenticate unencrypted Wi-Fi management frames at a single station without collaborative infrastructure (Jiang et al., 2012).

Later work generalized the same physical intuition to dynamic indoor scenarios. In mobile IoT settings, CSI pairs from adjacent times can be interpreted through temporal correlation for the legitimate device and distance-based correlation for an attacker, while in copresence systems the magnitude and phase across subcarriers serve as a wireless-context signature of shared physical surroundings. This suggests that CSI authentication is not a single algorithmic family, but a broader inference problem defined by what part of the signal is assumed to be stable, unique, or hard to forge in the target deployment (Guo et al., 28 Aug 2025, Fomichev et al., 2021).

2. Classical decision rules and protocol formulations

A basic formulation treats authentication as CSI similarity testing. In the mobile authentication scheme of "Practical Physical Layer Authentication for Mobile Scenarios Using a Synthetic Dataset Enhanced Deep Learning Approach" (Guo et al., 28 Aug 2025), the correlation-based baseline uses the Pearson correlation coefficient between CSI magnitude vectors and authenticates by comparing the resulting similarity score with an empirically chosen threshold. This establishes the canonical thresholding view: a legitimate pair should remain sufficiently correlated under expected channel evolution, whereas an attacker pair should not (Guo et al., 28 Aug 2025).

A more elaborate statistical rule appears in CSITE, which authenticates Wi-Fi management frames using a sliding window of recently trusted CSI samples. After dimensionality reduction to CSI amplitudes, the system computes a time-gain distance

$tgd(A,B) = dist(A,B)\cdot fc(A,B),$

where

$$dist(A,B) = \sqrt{\sum_{i=1}^{n}(A_i-B_i)^2}, \qquad fc(A,B) = e^{\lambda |t_A-t_B|}.$$

The degree of following is then defined as

$DoF(M) = \frac{1}{k}\sum_{i=1}^{k} tgd(M,P_i),$

and dynamic threshold scaling adapts the acceptance threshold to channel stability. In this formulation, CSI authentication is a continuously updated novelty test rather than a static one-shot comparison (Jiang et al., 2012).

Another protocol line embeds CSI in challenge-response authentication. "Hybrid Channel- and Coding-Based Challenge-Response Physical-Layer Authentication" (Crosara et al., 29 Jan 2025) assumes the receiver can partially control the channel, for example through an intelligent reflecting surface. The receiver selects channel configurations $\{\phi_k\}$, estimates the corresponding $h(\phi_k)$ from pilot symbols, and combines this CSI check with a wiretap-coded shared key. The stated security contribution is additive,

$b_{\text{hyb}} = b_{\text{ch}} + b_{\text{key}},$

and the core systems trade-off is between pilot allocation for CSI estimation and data allocation for coded key transmission (Crosara et al., 29 Jan 2025).

3. Deep learning architectures for CSI authentication

Deep learning entered CSI authentication first as an alternative to explicit channel modeling. "Supervised and Semi-Supervised Deep Neural Networks for CSI-Based Authentication" (Wang et al., 2018) introduced CNN, RNN, and CRNN architectures that learn directly from CSI. The CNN extracts local invariant features, the RNN models dependencies between frequencies, and the CRNN combines local and contextual information. In simulation, the reported accuracies were 96.9% for CNN-4 and 97.8% for CRNN-4, compared with 93.0% for SVM and 64.9% for the Neyman-Pearson test; in real USRP data, CNN, RNN, and CRNN reached 97.4%, 95.5%, and 98.1%, respectively. The same paper also proposed a semi-supervised pipeline in which K-means assigns pseudo-labels to unlabeled CSI before DNN pre-training and labeled fine-tuning (Wang et al., 2018).

A related but distinct objective is invariance to nuisance transformations. "Deep Neural Networks Meet CSI-Based Authentication" (Abyaneh et al., 2018) addressed the instability of correlation-based authentication under user rotation. Its LocNet architecture uses twin feature extractors to learn 10-dimensional features that remain stable toward rotation while still identifying location. The authentication protocol accepts a user if the fraction $\alpha$ of “same location” outputs exceeds a threshold $H_k = \mathfrak{Re}\{H_k\} + j\mathfrak{Im}\{H_k\}$0, and the reported apartment and garage experiments showed that valid locations were accepted despite device rotation while invalid locations were rejected (Abyaneh et al., 2018).

The most explicitly mobile deep-learning scheme in the current set of papers is the Siamese model of (Guo et al., 28 Aug 2025). There, a synthetic training dataset is generated from WLAN TGn models B–F, with temporal evolution governed by

$H_k = \mathfrak{Re}\{H_k\} + j\mathfrak{Im}\{H_k\}$1

and attacker CSI generated through a distance-correlation model. The authentication network uses twin CNN-based embedding branches, Euclidean distance in feature space, and a contrastive loss

$H_k = \mathfrak{Re}\{H_k\} + j\mathfrak{Im}\{H_k\}$2

with decision rule $H_k = \mathfrak{Re}\{H_k\} + j\mathfrak{Im}\{H_k\}$3 attacker. The synthetic data span TGn models B–F, SNR 5–50 dB, transmission intervals such as 3 ms, terminal speed such as 1 m/s, and attacker distance 0.25–3 wavelengths. On practical measurements, the reported AUC improvement is 0.03 over an FCN-based Siamese model and 0.06 over a correlation-based benchmark, while using significantly fewer parameters than the FCN-based Siamese model, 29,972 versus 282,641 (Guo et al., 28 Aug 2025).

4. Preprocessing, decomposition, and temporal consistency

A major theme in CSI authentication is that raw CSI mixes components useful for authentication with components useful for other security tasks, or with components that degrade stability. "On the Use of CSI for the Generation of RF Fingerprints and Secret Keys" (Srinivasan et al., 2021) and "Smart Channel State Information Pre-processing for Joint Authentication and Secret Key Distillation" (Srinivasan et al., 2022) both formalize a decomposition into predictable and unpredictable components. In these formulations, large-scale fading is a source of uniqueness and thus of RF fingerprinting, while small-scale fading is a source of shared entropy for secret key generation. PCA, kernel PCA, and autoencoders are used to separate these components; separability is quantified through total variation distance, and statistical independence through HSIC or dHSIC. A plausible implication is that CSI authentication performance depends not only on classifier design but also on whether the preprocessing stage preserves the component that is stable enough to authenticate and discards the component that is merely random (Srinivasan et al., 2021, Srinivasan et al., 2022).

Temporal consistency can also be enforced more directly. "Channel State Information Preprocessing for CSI-based Physical-Layer Authentication Using Reconciliation" (Passah et al., 18 Dec 2025) introduced A-RPCA, which jointly considers enrollment and authentication CSI and regularizes the low-rank component toward temporal consistency. The method is embedded in a reconciliation-based PLA framework using Gaussian approximation for Polar code design and short codelength Slepian-Wolf decoders. According to the abstract and detailed summary, A-RPCA substantially reduces the error probability after reconciliation, substantially increases the detection probabilities, and reaches detection probability 1 in both LOS and NLOS scenarios, outperforming baseline preprocessing schemes including PCA, RPCA, autoencoders, and ReProCS on synthetic and real datasets (Passah et al., 18 Dec 2025).

Signal-processing approaches without deep learning remain important, especially for replay resilience. "Wavelet-Based CSI Reconstruction for Improved Wireless Security Through Channel Reciprocity" (Basha et al., 10 Apr 2025) uses wavelet coherence and time-lagged cross-correlation to reconstruct aligned CSI traces between two devices. In its authentication protocol, the AP correlates its current CSI with the signed CSI returned by the STA and checks both correlation and time lag. The reported measurements show legitimate authentication at correlation approximately 0.8 with time shift approximately 1, versus replay attack at correlation approximately 0.04 with time shift approximately 330. This gives CSI temporal variation an explicitly freshness-oriented role: not merely identifying a sender, but detecting stale or replayed channel observations (Basha et al., 10 Apr 2025).

5. Extended forms: copresence, RF fingerprinting, biometrics, and open-set rejection

CSI authentication is not limited to source verification of a single transmitter. In copresence detection, the task is to decide whether two devices share the same local environment. "Next2You: Robust Copresence Detection Based on Channel State Information" (Fomichev et al., 2021) uses magnitude and phase from non-null subcarriers, feature-wise normalization, a fully connected neural network with four hidden layers, and temporal majority voting. It was evaluated on over 95 hours of CSI measurements across five real-world scenarios and reported error rates below 4%, while maintaining accurate detection in low-entropy context and insufficiently separated environments and supporting real-time operation on off-the-shelf smartphones (Fomichev et al., 2021).

A separate branch uses CSI-derived observables for RF fingerprinting. "DeepCSI: Rethinking Wi-Fi Radio Fingerprinting Through MU-MIMO CSI Feedback Deep Learning" (Meneghello et al., 2022) departs from explicit CSI extraction and instead authenticates MU-MIMO transmitters through standard-compliant beamforming feedback matrices. Because the feedback is collected during the sounding phase, it is stated to be unaffected by inter-user interference and inter-stream interference. The reported identification accuracy is up to 98%, remains above 82% when the device moves within the environment, and stays above 73% when training and testing positions do not perfectly overlap but are spatially well distributed (Meneghello et al., 2022).

Another fingerprinting line treats subtle CSI perturbations themselves as the fingerprint. "Physical-Layer Authentication of Commodity Wi-Fi Devices via Micro-Signals on CSI Curves" (Kong et al., 2023) introduced micro-CSI, attributed primarily to imperfections in RF circuitry, and combined a signal-space extraction algorithm with KNN matching to identify 11 COTS Wi-Fi NICs. The reported average attack detection rate is over 99% with a false alarm rate of 0%. "CSI-RFF: Leveraging Micro-Signals on CSI for RF Fingerprinting of Commodity WiFi" (Kong et al., 26 Feb 2026) extended this idea to open-set authentication, reporting that over a collection period extending beyond one year the extracted micro-CSI remained invariant over time, and that the authentication algorithm achieved an average attack detection rate close to 99% with a false alarm rate of 0% in both static and mobile conditions when using 20 CSI measurements to construct one fingerprint. "Towards Channel-Resilient CSI-Based RF Fingerprinting using Deep Learning" (Kong et al., 2024) then addressed channel variability explicitly through augmentation and supervised contrastive learning, reporting 97.13% accuracy at 40 dB SNR with a single CSI sample and improved robustness to previously unencountered channels. "Improving WiFi CSI Fingerprinting with IQ Samples" (Wang et al., 2024) pursued a different remedy: transforming CSI into time-domain signals in the IQ feature space and using auxiliary IQ training, improving recognition accuracy from 76% to 91% on a synthetic dataset and from 67% to 82% on a real dataset (Kong et al., 2023, Kong et al., 26 Feb 2026, Kong et al., 2024, Wang et al., 2024).

CSI has also been used as a biometric modality. "HandPass: A Wi-Fi CSI Palm Authentication Approach for Access Control" (Trindade et al., 25 Oct 2025) captured amplitude and phase from a Raspberry Pi-based setup and classified the right hands of 20 participants. With MinMax normalization and five evaluated classifiers, the Random Forest classifier achieved an average F1-Score of 99.82% using 10-fold cross-validation. In a different operational setting, "Dynamic and Open-Set RF Fingerprinting and Localization in Crowded Indoor Environments through Contrastive Channel State Information Learning" (Razak et al., 15 May 2026) proposed ContraCSI, a contrastive learning framework using ESP32 devices. For closed-set identification, ViT variants gave the best overall performance, while open-set authentication used a GEM-based anomaly score and sequential CUSUM on embeddings learned by Lite3D-CNN-Contra, with reported average open-set F1 and AUC of 0.98 over several unseen transmitter cases (Trindade et al., 25 Oct 2025, Razak et al., 15 May 2026).

6. Security evaluation, limitations, and emerging directions

Recent synthesis work argues that CSI authentication has often been evaluated too narrowly. "SoK: Security Evaluation of Wi-Fi CSI Biometrics: Attacks, Metrics, and Systemic Weaknesses" (Braga et al., 14 Nov 2025) identifies replay attacks, geometric mimicry attacks, and environmental perturbation attacks, and argues that aggregate accuracy alone is insufficient. It recommends reporting FAR, FRR, EER, per-class EER, Frequency Count of Scores, and the Gini Coefficient, and stresses that high average performance can mask risk concentration in particular users or classes. The same work concludes that CSI-based authentication should be considered complementary, not a sole authentication primitive, unless it is evaluated under realistic attack models and with per-user risk analysis (Braga et al., 14 Nov 2025).

The literature also documents practical constraints internal to specific methods. The mobile Siamese approach of (Guo et al., 28 Aug 2025) depends on accurate channel modeling, empirical threshold selection, and coverage of the relevant SNR, delay-spread, and attacker-distance ranges in the synthetic dataset. Micro-CSI extraction in (Kong et al., 2023) is explicitly developed for line-of-sight scenarios, and the reported performance dips when attacker and legitimate devices are of the same model. HandPass reports its strongest results in a controlled LOS setting with standardized hand placement, reduced antenna power, and removal of jewelry, while spoofing and adversarial attacks are not directly evaluated. These results do not invalidate CSI authentication, but they bound the claims that can be made about portability across environments, hardware, and adversarial conditions (Guo et al., 28 Aug 2025, Kong et al., 2023, Trindade et al., 25 Oct 2025).

An emerging direction is adaptive generation or extrapolation of CSI fingerprints. "APEG: Adaptive Physical Layer Authentication with Channel Extrapolation and Generative AI" (Cheng et al., 23 Mar 2026) uses collaborator-assisted channel extrapolation and diffusion models, specifically CCMDM and CADM, to generate expected CSI fingerprints in dynamic scenarios. According to the reported simulation results, CADM achieves F1 = 1.0 and error rate $H_k = \mathfrak{Re}\{H_k\} + j\mathfrak{Im}\{H_k\}$4 from 5 dB to 20 dB SNR, while CCMDM shows a significant advantage in convergence speed. This suggests a shift from direct similarity matching toward model-based prediction of what the legitimate CSI should be under current conditions, although the evidence reported here is simulation-based rather than a deployment study (Cheng et al., 23 Mar 2026).