pith. sign in

arxiv: 2604.17192 · v1 · submitted 2026-04-19 · 📡 eess.SP

CADRE: Card-Agnostic Domain-Aligned RF Embeddings for Virtual PIN Pads on Passive NFC Cards

Dickson Akuoko Sarpong , Hongzhi Guo This is my paper

Pith reviewed 2026-05-10 06:33 UTC · model grok-4.3

classification 📡 eess.SP
keywords NFCvirtual PIN padRF embeddingsdomain alignmentadversarial trainingMahalanobis distancepassive cardsimpedance variation
0
0 comments X

The pith

Adversarial alignment of RF embeddings from card presses turns impedance variations into card-invariant features for consistent PIN recognition.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes a framework for adding password security to passive NFC cards by letting users press designated regions that alter the card's antenna impedance and produce detectable changes in the reader's RF signals. A temporal neural encoder extracts features from these signals, after which an adversarial module aligns the embeddings at the penultimate layer so that identical presses yield similar representations regardless of which commercial card is used. Recognition then proceeds with a Mahalanobis distance classifier trained on a covariance model derived from calibration data. If this alignment succeeds, existing NFC readers can authenticate users without any modification to the passive cards themselves. The reported outcome is a 98.20 percent acceptance rate that holds under added noise.

Core claim

The authors claim that an adversarial domain-alignment module applied directly to the RF feature space at the penultimate layer of a temporal neural encoder can reshape press-induced signal embeddings into compact, card-invariant clusters. When these clusters are classified using a Mahalanobis distance metric obtained from a calibration-based covariance model, the resulting system achieves a 98.20 percent recognition acceptance rate across heterogeneous ISO/IEC 15693 cards and remains stable under substantial noise degradation.

What carries the argument

adversarial domain-alignment module that reshapes press-response RF embeddings into compact, card-invariant clusters at the penultimate layer of a temporal neural encoder, followed by Mahalanobis-distance classification

Load-bearing premise

That press-induced impedance variations produce signal patterns distinguishable enough to be aligned into card-invariant clusters by adversarial training at the penultimate layer, enabling reliable Mahalanobis classification across different commercial cards.

What would settle it

Evaluating the system on a fresh collection of NFC cards from additional manufacturers and finding that the recognition acceptance rate falls substantially below 90 percent would falsify the claim of reliable card-agnostic performance.

Figures

Figures reproduced from arXiv: 2604.17192 by Dickson Akuoko Sarpong, Hongzhi Guo.

Figure 1
Figure 1. Figure 1: Application example of the virtual PIN pad in NFC door mounted access system. and displays to allow password input after the card is tapped. However, this approach increases the cost and complexity of the reader. Additionally, in certain scenarios, such as during a pandemic, users may be unwilling to touch shared PIN pads. Prior work has also shown that passwords entered on reader-side PIN pads can be infe… view at source ↗
Figure 2
Figure 2. Figure 2: A simplified circuit diagram for the NFC reader-card [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Buttons predefined spatial positions on an NFC card. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of button coil-induced coupling on reader current [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: COMSOL Multiphysics simulation of (a) the spatial [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 9
Figure 9. Figure 9: Response signals across 9 button positions on a card. [PITH_FULL_IMAGE:figures/full_fig_p005_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Data collection experiment setup. origin, the button coordinates are x ∈ {−27.5, 0, +27.5}mm and y ∈ {+18, 0, −18}mm. This results in the but￾ton centers as follows, Button 0: (−27.5, +18)mm, But￾ton 1: (0, +18)mm, Button 2: (+27.5, +18)mm, Button 3: (−27.5, 0)mm, Button 4: (0, 0)mm, Button 5: (+27.5, 0)mm, Button 6: (−27.5, −18)mm, Button 7: (0, −18)mm, and But￾ton 8: (+27.5, −18)mm. For each button posit… view at source ↗
Figure 10
Figure 10. Figure 10: Adversarial domain adaptation architecture with [PITH_FULL_IMAGE:figures/full_fig_p006_10.png] view at source ↗
Figure 13
Figure 13. Figure 13: Illustration of Mahalanobis distance concept. [PITH_FULL_IMAGE:figures/full_fig_p007_13.png] view at source ↗
Figure 12
Figure 12. Figure 12: Confusion matrices for button class prediction using [PITH_FULL_IMAGE:figures/full_fig_p007_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: Impact of risk levels (α) on FAR and FRR. obtaining the whitening transform via the Cholesky factor￾ization of the d × d covariance matrix incurs O(d 3 ). During inference, evaluating the Mahalanobis distance for each class involves solving a triangular system followed by a squared norm computation, giving a per button class cost of O(d 2 ) [19]. The overall time complexity for recognizing a single test s… view at source ↗
Figure 15
Figure 15. Figure 15: Average per-button FAR and FRR for unseen cards. [PITH_FULL_IMAGE:figures/full_fig_p009_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Recognition performance using different methods. [PITH_FULL_IMAGE:figures/full_fig_p009_16.png] view at source ↗
read the original abstract

Near Field Communication (NFC) cards are widely used for identification, but their passive nature often limits the ability to incorporate additional security mechanisms. As a result, anyone holding the card may be incorrectly recognized as an authenticated user. To overcome this limitation, this paper presents a secure manual password input framework using a virtual PIN pad for passive NFC cards. Users input passwords by pressing designated regions on the card, which induces measurable impedance variations in the NFC antenna. These variations change the RF signals subtly, and a deep learning model is used to infer the intended password from the resulting signal patterns. A key challenge is that identical press interactions can produce significantly different responses across NFC cards, which yields unreliable recognition. To address this, we introduce a lightweight recognition approach that operates directly within the RF feature space at the penultimate layer of a temporal neural encoder. An adversarial domain-alignment module reshapes virtual PIN pad press-response embeddings into compact, card-invariant clusters, which enables stable and consistent recognition across heterogeneous cards. To support model training and evaluation, a reconfigurable software-defined radio (SDR) testbed is developed, and PIN pad press-response data are collected from commercially available ISO/IEC 15693 cards. Recognition is performed using a Mahalanobis distance metric derived from a calibration-based covariance model that captures feature correlations. Experimental results show that the proposed system achieves a 98.20\% recognition acceptance rate and remains robust under substantial noise degradation. The framework is fully card-agnostic and can be seamlessly integrated into existing NFC infrastructures.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes CADRE, a card-agnostic framework for secure virtual PIN pad input on passive NFC cards. Users press designated regions on ISO/IEC 15693 cards, inducing impedance variations that alter RF signals; a temporal neural encoder extracts features, an adversarial domain-alignment module at the penultimate layer produces card-invariant embeddings, and Mahalanobis-distance classification (with calibration-based covariance) decodes the password. The abstract reports 98.20% recognition acceptance rate with robustness to noise and seamless integration into existing NFC infrastructures.

Significance. If the card-agnostic property and reported accuracy hold under proper cross-card validation, the work would enable password-based authentication on unmodified passive NFC cards, addressing a practical security gap in contactless identification systems. The use of SDR-collected real-card data and adversarial alignment at the feature level is a concrete technical contribution, though its generalizability remains to be demonstrated.

major comments (3)
  1. Abstract: The 98.20% recognition acceptance rate and noise-robustness claims are presented without any dataset size, number of distinct commercial card models, cross-card hold-out protocol, or ablation studies on the adversarial module. This directly undermines the central claim that the system is fully card-agnostic and works across heterogeneous cards.
  2. Abstract and experimental description: No evidence is supplied that test cards were excluded from the adversarial alignment training or calibration covariance estimation. If alignment was performed on the same cards used for testing, the reported performance does not establish generalization to unseen commercial cards, which is load-bearing for the 'card-agnostic' assertion.
  3. Abstract: The Mahalanobis-distance classifier relies on a calibration-based covariance model, yet no details are given on how many presses per card or per region were used for covariance estimation, nor on the stability of the resulting clusters when new card hardware is introduced.
minor comments (2)
  1. Abstract: The phrase 'substantial noise degradation' is vague; quantitative SNR values or noise models used in the robustness experiments should be stated.
  2. The manuscript should include a clear diagram or pseudocode for the adversarial domain-alignment loss and the precise location of the penultimate layer within the temporal encoder.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our experimental validation. We will revise the manuscript to provide the requested details on dataset characteristics, validation protocols, and calibration procedures, thereby strengthening support for the card-agnostic claims.

read point-by-point responses

  1. Referee: Abstract: The 98.20% recognition acceptance rate and noise-robustness claims are presented without any dataset size, number of distinct commercial card models, cross-card hold-out protocol, or ablation studies on the adversarial module. This directly undermines the central claim that the system is fully card-agnostic and works across heterogeneous cards.

    Authors: We agree that the abstract should supply more context on the experimental setup. In the revised manuscript we will update the abstract to include the dataset size, the number of distinct commercial card models, the cross-card hold-out protocol, and a reference to the ablation studies on the adversarial module. We will also expand the experimental section to present these elements explicitly. revision: yes

  2. Referee: Abstract and experimental description: No evidence is supplied that test cards were excluded from the adversarial alignment training or calibration covariance estimation. If alignment was performed on the same cards used for testing, the reported performance does not establish generalization to unseen commercial cards, which is load-bearing for the 'card-agnostic' assertion.

    Authors: We will revise both the abstract and the experimental description to explicitly state that test cards were held out from adversarial alignment training and from covariance estimation. The revised text will describe the cross-card hold-out protocol in detail and report the associated performance to demonstrate generalization to unseen cards. revision: yes

  3. Referee: Abstract: The Mahalanobis-distance classifier relies on a calibration-based covariance model, yet no details are given on how many presses per card or per region were used for covariance estimation, nor on the stability of the resulting clusters when new card hardware is introduced.

    Authors: We will add the missing details on the number of presses per card and per region used for covariance estimation. We will also include an analysis of cluster stability when new card hardware is introduced, either through additional observations from the existing multi-card dataset or targeted supplementary experiments. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper describes an empirical ML pipeline for classifying press-induced RF variations on NFC cards via a temporal encoder plus adversarial domain alignment at the penultimate layer, followed by Mahalanobis classification on collected SDR data. No equations, derivations, or claims reduce by construction to fitted parameters or self-referential definitions; recognition rates are reported directly from independent test collections on real ISO/IEC 15693 cards. No self-citations are invoked as load-bearing uniqueness theorems or to smuggle ansatzes, and the central claims rest on standard adversarial training and covariance modeling without renaming known results or importing prior author work as external fact. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that RF signal variations from finger presses are sufficiently informative and alignable across cards; no free parameters or new physical entities are explicitly introduced in the abstract.

axioms (2)
  • domain assumption Temporal neural encoders can extract press-specific features from subtle RF impedance variations induced by finger contact on NFC antennas.
    Invoked by the use of the penultimate layer embeddings and adversarial alignment module.
  • domain assumption Adversarial domain alignment can produce compact card-invariant clusters without destroying press-discriminative information.
    Core to the lightweight recognition approach described.

pith-pipeline@v0.9.0 · 5582 in / 1400 out tokens · 52642 ms · 2026-05-10T06:33:30.431880+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

  1. [1]

    Idcc: Influence-driven content cache for nfc in ioe,

    R. Wang, Y . Shen, W. Wan, B. Yue, S. Wu, and Y . Zhang, “Idcc: Influence-driven content cache for nfc in ioe,”IEEE Internet of Things Journal, vol. 12, pp. 28 319–28 331, 2025

    work page 2025

  2. [2]

    Model-agnostic uncertainty quantification for fast nfc tag identification using rf finger- printing,

    D. A. Sarpong, A. Kamrath, R. Bhusal, and H. Guo, “Model-agnostic uncertainty quantification for fast nfc tag identification using rf finger- printing,”IEEE Internet of Things Journal, vol. 12, pp. 47 607–47 622, 2025

    work page 2025

  3. [3]

    Enhanced emv security: Preventing credit card fraud from a distance,

    M.-H. Yang, Y .-S. Hsu, and H.-C. Hsu, “Enhanced emv security: Preventing credit card fraud from a distance,”IEEE Access, vol. 13, pp. 88 858–88 870, 2025

    work page 2025

  4. [4]

    Dynamic precoding for near-field secure communications: Implementation and performance analysis,

    Z. Teng, J. An, C. Masouros, H. Li, L. Gan, and D. W. K. Ng, “Dynamic precoding for near-field secure communications: Implementation and performance analysis,”IEEE Internet of Things Journal, vol. 12, pp. 29 427–29 442, 2025

    work page 2025

  5. [5]

    Dcnn-rff-nfc: a novel design of nfc security using deep convolution neural network-based rf finger- printing,

    S. Adigopula and M. V . Subramanyam, “Dcnn-rff-nfc: a novel design of nfc security using deep convolution neural network-based rf finger- printing,”Neural Computing and Applications, vol. 37, pp. 4439 – 4453, 2024

    work page 2024

  6. [6]

    Jump out of resonance: A practical nfc tag fingerprinting scheme,

    Y . Yang, Z. An, J. Cao, Y . Wang, P. Hu, G. Zhang, and X. Cheng, “Jump out of resonance: A practical nfc tag fingerprinting scheme,” IEEE Transactions on Mobile Computing, vol. 23, pp. 8694–8709, 2024

    work page 2024

  7. [7]

    Communication and power transfer analy- sis of interfering magnetically resonant coupled systems,

    R. Fischbacher, J. R. Lopera, D. J. Pommerenke, R. Prestros, B. Auinger, W. Bösch, and J. Grosinger, “Communication and power transfer analy- sis of interfering magnetically resonant coupled systems,”IEEE Journal of Radio Frequency Identification, vol. 8, pp. 713–723, 2024

    work page 2024

  8. [8]

    Inductance calculations for noncoaxial coils using bessel functions,

    J. T. Conway, “Inductance calculations for noncoaxial coils using bessel functions,”IEEE Transactions on Magnetics, vol. 43, no. 3, pp. 1023– 1034, 2007

    work page 2007

  9. [9]

    Finkenzeller,RFID Handbook: Fundamentals and Applications in Contactless Smart Cards and Identification, 2nd ed

    K. Finkenzeller,RFID Handbook: Fundamentals and Applications in Contactless Smart Cards and Identification, 2nd ed. Wiley Publishing, 2003

    work page 2003

  10. [10]

    M2i communication: From theoretical modeling to practical design,

    H. Guo and Z. Sun, “M2i communication: From theoretical modeling to practical design,”2016 IEEE International Conference on Communi- cations (ICC), pp. 1–6, 2015

    work page 2016

  11. [11]

    Identification cards — Contactless inte- grated circuit cards — Vicinity cards,

    ISO, Geneva, Switzerland, “Identification cards — Contactless inte- grated circuit cards — Vicinity cards,” no. ISO 15693, 2006

    work page 2006

  12. [12]

    ICODE SLIX – SL2S2002, ISO/IEC 15693 compliant smart label IC,

    NXP Semiconductors, “ICODE SLIX – SL2S2002, ISO/IEC 15693 compliant smart label IC,” Product Data Sheet, Rev. 3.7, Dec 2018

    work page 2018

  13. [13]

    Unsupervised domain adapta- tion for wifi gesture recognition,

    B. Zhang, D. Zhang, Y . Hu, and Y . Chen, “Unsupervised domain adapta- tion for wifi gesture recognition,”2023 IEEE Wireless Communications and Networking Conference (WCNC), pp. 1–6, 2023

    work page 2023

  14. [14]

    Domain neural adapta- tion,

    S. Chen, Z. Hong, M. Harandi, and X. Yang, “Domain neural adapta- tion,”IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 11, pp. 8630–8641, 2023

    work page 2023

  15. [15]

    In or out? fixing imagenet out-of-distribution detection evaluation,

    J. Bitterwolf, M. Müller, and M. Hein, “In or out? fixing imagenet out-of-distribution detection evaluation,” inProceedings of the 40th International Conference on Machine Learning, ser. ICML, 2023

    work page 2023

  16. [16]

    Efficient implementation of mahalanobis distance on ferroelectric finfet crossbar for outlier detection,

    M. Rafiq, Y . S. Chauhan, and S. Sahay, “Efficient implementation of mahalanobis distance on ferroelectric finfet crossbar for outlier detection,”IEEE Journal of the Electron Devices Society, vol. 12, pp. 516–524, 2024

    work page 2024

  17. [17]

    Diversify: A general framework for time series out-of-distribution detection and generalization,

    W. Lu, J. Wang, X. Sun, Y . Chen, X. Ji, Q. Yang, and X. Xie, “Diversify: A general framework for time series out-of-distribution detection and generalization,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, pp. 4534–4550, 2023

    work page 2023

  18. [18]

    Polynomial whitening for high-dimensional data,

    J. Gillard, E. O’Riordan, and A. Zhigljavsky, “Polynomial whitening for high-dimensional data,”Comput. Stat., vol. 38, no. 3, p. 1427–1461, Sep. 2022

    work page 2022

  19. [19]

    Anomaly detection using minimum covariant determinant as feature in multivariate data,

    A. Shrivastava and P. R. Vamsi, “Anomaly detection using minimum covariant determinant as feature in multivariate data,”Proceedings of the 2023 Fifteenth International Conference on Contemporary Computing, pp. 501–514, 2023

    work page 2023

  20. [20]

    Gaus- sian latent representations for uncertainty estimation using mahalanobis distance in deep classifiers,

    A. Venkataramanan, A. Benbihi, M. Laviale, and C. Pradalier, “Gaus- sian latent representations for uncertainty estimation using mahalanobis distance in deep classifiers,”2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pp. 4490–4499, 2023

    work page 2023