Pith. sign in

REVIEW 6 major objections 5 minor 36 references

This paper claims that a framework called DRIFT, which explicitly disentangles transmitter-specific from receiver-specific representations in received I/Q signals, lets an RF fingerprint identification model trained on a few known receivers

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

DRIFT disentangles transmitter and receiver features and uses adversarial alignment plus receiver-clustering regularization to keep RF fingerprint identification accurate on unseen receivers.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A genuinely useful cross-receiver RFFI method with a consistent empirical win, but the dataset description doesn't match the experiments and the theory is more assertion than proof. the 6 major comments →

arxiv 2510.09405 v2 pith:UU22RKNE submitted 2025-10-10 cs.LG

Cross-Receiver Generalization for RF Fingerprint Identification via Feature Disentanglement and Adversarial Training

Yuhao Pan , Xiucheng Wang , Fushuo Huo , Nan Cheng , Wenchao Xu This is my paper
classification cs.LG
keywords RF fingerprint identificationcross-receiver generalizationfeature disentanglementadversarial traininggradient reversal layerdomain generalizationreceiver-invariant representationWiSig ManySig
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper is trying to establish that receiver-induced hardware variability, which wrecks radio frequency fingerprint identification when the receiving device is swapped, can be handled by separating the learned signal representation into two halves: one that carries transmitter identity and one that carries receiver style. It argues that a multiplicative signal model, X = g_j · h_ij · f_i · s_i(t), lets a neural network learn which features belong to the transmitter and which belong to the receiver, and that adversarial alignment plus a center-based clustering loss can strip receiver information out of the transmitter features. If correct, this means a model trained on three to five known receivers can be dropped onto a new receiver without retraining or target-domain data, a practically important capability for IoT device authentication. The experiments on the WiSig ManySig dataset support the claim: DRIFT reaches average cross-receiver accuracies of 74–83%, about 2 to 14 points above existing domain-generalization baselines.

Core claim

The central claim is that cross-receiver generalization does not require seeing the new receiver during training; it only requires learning a representation in which receiver-specific impairments occupy a separable, removable subspace. The paper proposes DRIFT, which takes the 512-dimensional embedding from a ResNet-18 backbone and splits it in half: the first 256 dimensions are trained to be transmitter-discriminative and receiver-invariant via a gradient reversal layer, while the second 256 dimensions are trained to predict the receiver identity and pulled toward a per-receiver center. A negative MSE loss pushes the two halves apart in latent space. Under the assumption in Eq. (1) that rec

What carries the argument

The central object is the multiplicative signal model X_t_ij = g_j · h_t_ij · f_i · s_i(t), which attributes receiver-specific distortion g_j separately from transmitter-specific hardware impairments f_i. The machinery that carries the argument is the fixed dimension-wise split in Eq. (17): the 512-dimensional embedding is split into a 256-dim transmitter half and a 256-dim receiver half. Three losses act on these halves: a gradient reversal layer (GRL) adversarially aligns the transmitter half across receivers, a center loss clusters the receiver half by receiver identity, and a negative-MSE separation loss maximizes the distance between the two halves. Together these enforce the disentangl

Load-bearing premise

The entire argument rests on the premise that receiver-induced distortion is a multiplicative, separable factor g_j that occupies a fixed, distinct half of the learned feature space, so that removing that half and aligning the rest preserves transmitter identity; if receiver distortion is nonlinear, frequency-dependent, or entangled with transmitter features in a way that violates Eq. (1), the cross-receiver guarantee does not follow.

What would settle it

Train DRIFT on several receivers as in the paper, then use a receiver whose distortion is deliberately manufactured to violate the multiplicative model—for example, a receiver with a strong frequency-selective filter or a nonlinear memory effect—and measure accuracy on that unseen receiver; if accuracy drops to the level of the basic ERM baseline (around 60–65%), the separable-subspace assumption is not general. Alternatively, after training, freeze the feature extractor and train a receiver classifier on the transmitter-specific half z*; if that classifier can predict the receiver with well a

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If DRIFT is correct, a deployed RFFI model can be replaced by a new receiver without collecting labeled data from that receiver, and still maintain roughly 74–83% identification accuracy.
  • Training on three to five receivers with the same set of transmitters is sufficient to learn receiver-invariant transmitter features, making the approach practical for IoT networks with a small number of known receiver nodes.
  • The simplified generalization bound (Eq. 14) provides a principled target: lowering pairwise H-divergence between source domains and the distance to the target's convex hull reduces worst-case unseen-receiver risk.
  • The ablation study in Table III shows that all three modules (GRL, center loss, MSE separation) are complementary; removing any combination degrades accuracy, indicating that the disentanglement mechanism, not any single loss, drives the gains.
  • The cross-day experiments (Table II) suggest the same disentangled features are also robust to temporal channel variation, not just receiver replacement, extending the method's practical reach.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's fixed, hand-assigned split of the embedding into transmitter and receiver halves is a strong inductive bias; a learned or adaptive partition might generalize better to receivers whose impairments do not line up with the first or second 256 dimensions.
  • The same disentanglement recipe could be extended to channel effects as a third factor, separating h_t_ij from both transmitter and receiver; if it works, it would yield a channel-, receiver-, and time-agnostic fingerprinting system.
  • A testable extension is to evaluate DRIFT on receivers with strongly nonlinear or frequency-selective front-ends (e.g., heavy clipping or severe I/Q imbalance) to see whether the multiplicative model in Eq. (1) still holds; a collapse in accuracy would pinpoint the boundary of the assumption.
  • The center loss on receiver features implicitly assumes each receiver has a single style; a receiver with mode-dependent distortions (e.g., temperature-dependent clock drift) might need a mixture-of-centers variant.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

6 major / 5 minor

Summary. The paper proposes DRIFT, a cross-receiver domain-generalization framework for RF fingerprint identification. The signal model in Eq. (1) factors a received I/Q sample into transmitter hardware impairments, channel response, and receiver hardware impairments. DRIFT first extracts a 512-d representation with a 1D ResNet-18, splits it by dimension into a 256-d transmitter-specific and 256-d receiver-specific part, and then applies a transmitter classifier, a receiver classifier, a GRL adversarial receiver discriminator on the transmitter part, a center loss on the receiver part, and a negative-MSE separation loss. A domain-generalization bound is invoked to motivate reducing the H-divergence terms. Experiments on a WiSig subset compare DRIFT against DANN, RIEI, MTL, and ERM under multiple training receiver sets, unseen test receivers, and cross-day evaluation. The paper reports consistent improvements, with average accuracies from 74% to 83% versus 60–81% for baselines.

Significance. Cross-receiver RFFI is a practically important problem, and the reported gains, if correct, would substantively improve receiver-agnostic device identification. The paper includes a physically motivated signal model, public-dataset experiments, four baselines, ablations of each loss term, and a hyperparameter study. These are useful contributions. However, the immediate significance is limited by the absence of code or data-split release, the lack of multiple-seed variance reporting, and the unresolved dataset-composition inconsistency that affects the benchmark identity. The theory section also asserts rather than derives the key mechanism. The empirical ranking of DRIFT is consistently favorable across tables, which supports the core claim more than any individual number, but the above issues must be resolved before the result can be accepted as stated.

major comments (6)
  1. [Section VI.A and VI.C, Table I] The dataset composition is internally inconsistent. Section VI.A states that the ManySig subset 'consists of six receivers and twelve transmitters,' yet the experiments use twelve distinct receiver IDs (1-1, 1-19, 14-7, 18-2, 7-7, 8-8, 19-2, 2-1, 2-19, 20-1, 7-14, 3-19). The training-sample formula '6×800×n' and test formula '6×200' imply six transmitters, not twelve. This is exactly the composition of the ManyRx subset rather than ManySig. The mismatch changes the label count K, the number of domains, and the benchmark identity, so the reported averages and the 'consistently outperforms' conclusion are scoped to a benchmark different from the one described. The authors must clarify which subset was actually used and either rerun on the claimed subset or correct the dataset description, and provide the exact Tx/Rx split.
  2. [Section IV.B, after Eq. (14)] The theoretical argument asserts, without proof, that disentangling transmitter and receiver features reduces gamma and epsilon. Specifically, the text states that pairwise source divergence and the target-to-mixture divergence 'can be reduced' after separation, but no derivation, theorem, or empirical measurement is given. A GRL, center loss, or dimension split can in principle also increase representation divergence. Either prove the reduction under the assumptions of Eq. (1) and Eq. (17), or reframe this section as heuristic motivation and support it with measured gamma/epsilon values or ablation evidence.
  3. [Section V.A, Eq. (17)] The central mechanism assumes that receiver-specific impairments occupy exactly the second half of a 512-d embedding and that transmitter-specific features occupy the first half. This fixed dimension split is an unvalidated architectural assumption. There is no evidence that receiver distortion is linearly separable in this coordinate-aligned subspace, or that GRL and center loss remove all receiver information that does not fall in that half. The paper should test this assumption, e.g., by varying the split ratio, visualizing the separated subspaces, or measuring residual receiver information in z*.
  4. [Section VI.F and Table I] The hyperparameters lambda_1=1, lambda_2=0.01, lambda_3=0.02 were selected using a sensitivity analysis performed on the training receiver combination {1-1, 14-7, 18-2, 7-7}, and that same combination is a headline configuration in Table I and Table II. Tuning on a member of the evaluation family can inflate the reported result for that configuration. The default hyperparameters should be fixed on a separate validation split, or results for all configurations should also be reported with a single default hyperparameter set that was not tuned on any tested configuration.
  5. [Algorithm 1 line 11 and Section V.C Eqs. (23)-(24)] The gradient update has a sign inconsistency in the adversarial component. Eq. (24) says the GRL backward pass multiplies the gradient by -lambda I. In Algorithm 1, the update is written as theta <- theta - eta (nabla L_CE - lambda_1 nabla L_grl + ...). If the GRL already reverses the gradient, this subtracts the adversarial gradient a second time; if the minus sign in the update is meant to implement the reversal, then Eq. (22)-(24) and the forward/backward definition of GRL are not stated consistently. The exact PyTorch/autograd implementation of the GRL stop-gradient and the sign convention in the total loss need to be clarified for reproduction.
  6. [All experiments, especially Tables I-III] The experiments use a fixed random seed and average only the last five epochs. No error bars, multiple seeds, or statistical significance tests are reported. Since some headline differences are only a few percentage points (e.g., DRIFT 79.27 vs. RIEI 71.53 in Table II Group 1, or DRIFT 69.67 vs. RIEI 69.37 in Group 2), the claim of consistent superiority needs variance estimates or at least three seeds with standard deviations. The fix-seed statement in Section VI.B does not replace this.
minor comments (5)
  1. [Section VI.C, first paragraph] The text reports DANN's four-receiver average as 62.68%, but Table I shows DANN average 60.85% for the {1-1, 14-7, 18-2, 7-7} training set. The MTL average is also omitted in that sentence.
  2. [Section IV.B, last sentence] 'The detailed design of the proposed framework is presented in Section IV' should refer to Section V.
  3. [Section VI.B, Algorithm 2] The method named DANN is attributed to Shen et al. [14]. Domain-adversarial neural networks originate with Ganin et al. [36]; Shen et al. build on this idea. The citation should distinguish the original DANN from the receiver-GRL variant used here.
  4. [Table II] The test receivers for the cross-day experiments are not enumerated, and only averaged accuracies are shown. List the exact test receiver set for each training group and ideally report per-receiver, per-day results to make the cross-day claim independently assessable.
  5. [Algorithm 1, line 5] The center c_d is updated inside the mini-batch loop, but the notation 1/|S_d| sum over S_d suggests the full receiver domain. Clarify whether the center is computed on the current batch or maintained as a global moving average.

Circularity Check

0 steps flagged

No significant circularity: DRIFT is validated against external baselines on a public dataset, and its theory is a standard DG bound with a stated assumption, not an input–output identity.

full rationale

The paper's central empirical claim—that DRIFT 'consistently outperforms state-of-the-art baselines under cross-receiver evaluation'—is established by comparing against independently published methods (DANN, RIEI, MTL, ERM) on the public WiSig dataset. These baselines are not the authors' own prior results, so the comparison has independent content. The theoretical Section IV imports a standard domain-generalization bound from prior work [32]–[34], then simplifies it under the explicit assumption in Eq. (10) that source and target label functions coincide; this is a stated modeling assumption, not a hidden use of the target result. The subsequent claim that disentanglement reduces γ and ε is an informal motivation rather than a circular derivation: the paper does not fit those quantities from test data and then present them as predictions. The only self-citations are unrelated radio-map papers [11], [12] and are not load-bearing. Two concerns noted by the reader are not circularity: the hyperparameter sensitivity analysis in Section VI.F selects default weights on one of the reported receiver combinations, which is a model-selection caveat, and the dataset-composition inconsistency between Section VI.A ('six receivers and twelve transmitters') and the training-sample formula '6×800×n' in Section VI.C is a reproducibility issue. Neither reduces a claimed result to its own input by construction. Hence the paper is not circular as defined by the rubric.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central result rests on a set of modeling choices: the multiplicative channel/receiver/transmitter signal model (Eq. 1), identical label functions across domains (Eq. 10), the clustering assumption behind L_center, the cited generalization bound, and the hard-coded 256/256 feature split. The loss weights and other hyperparameters are chosen by hand or sensitivity analysis. No new physical entities are introduced.

free parameters (7)
  • lambda_1 (GRL loss weight) = 1
    Chosen by hyperparameter sensitivity on one train/test split; controls adversarial alignment strength.
  • lambda_2 (center loss weight) = 0.01
    Chosen by hyperparameter sweep; the paper notes performance is highly sensitive to this value.
  • lambda_3 (negative MSE separation weight) = 0.02
    Chosen by hyperparameter sweep; reported to be less sensitive.
  • feature split ratio = 256/256 (1:1)
    Hard-coded partition of the 512-d embedding into transmitter/receiver halves; no experimentation or justification for the split dimension.
  • training epochs T = not stated
    Algorithm 1 takes T as input but the value is never reported; checkpoint selection from the last five epochs depends on it.
  • learning rate eta = 0.0001
    Standard Adam learning rate chosen by hand; affects convergence and final accuracy.
  • batch size B = 64
    Mini-batch size used in all experiments; chosen by hand.
axioms (6)
  • domain assumption Received I/Q signal factorizes as X = g_j * h_ij * f_i * s_i(t) (Eq. 1).
    Assumes receiver-specific distortion is a multiplicative, separable term; if not, fixed subspace disentanglement cannot isolate it.
  • domain assumption All source and target domains share the same label function f_S1 = ... = f_Sn = f_U (Eq. 10).
    Used to drop the label-discrepancy term from Theorem 1 and to allow transmitter labels to transfer to unseen receivers.
  • domain assumption Signals from different transmitters received by one receiver share similar receiver characteristics, justifying clustering z' around receiver centroids (Eq. 25).
    Load-bearing premise for the L_center regularization.
  • standard math Standard domain generalization bound (Theorem 1) from [32]-[34] is valid for this setting.
    The paper's theoretical argument is built on this cited bound.
  • domain assumption Gradient reversal (Eqs. 23-24) makes features domain-invariant in the presence of a finite discriminator.
    Assumes adversarial alignment removes residual receiver information from z*.
  • ad hoc to paper The 512-d embedding can be partitioned into orthogonal transmitter and receiver subspaces by dimension index.
    No evidence is provided that receiver and transmitter features live in disjoint dimensions; this is the central architectural bet.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Receiver Generalization for RF Fingerprint Identification via Feature Disentanglement and Adversarial Training." pith.science (2026). https://pith.science/paper/UU22RKNE

@misc{pith2026251009405,
  author       = {Pith},
  title        = {Pith review of: Cross-Receiver Generalization for RF Fingerprint Identification via Feature Disentanglement and Adversarial Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UU22RKNE}},
  note         = {Machine review of arXiv:2510.09405}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Radio frequency fingerprint identification (RFFI) is a key technique for wireless network security, leveraging intrinsic hardware imperfections to enable transmitter identification. Although deep neural networks are effective at extracting discriminative RF features, their performance is significantly affected by receiver-induced variability in practical deployments. In real-world scenarios, RF signals inherently entangle transmitter-specific characteristics with receiver-dependent distortions, leading models to capture receiver-related patterns when training and evaluation are conducted on the same device. Consequently, replacing the receiver during deployment often results in notable performance degradation. To address this issue, we propose a cross-receiver robust RFFI framework that explicitly disentangles transmitter-specific and receiver-specific representations. The proposed method integrates adversarial domain alignment with receiver-aware regularization to suppress residual receiver information in transmitter features while enforcing intra-receiver consistency in receiver-specific representations. A feature separation constraint is further introduced to decouple the two components in the latent space. Extensive experiments on multi-receiver WiFi datasets demonstrate that the proposed method consistently outperforms state-of-the-art baselines under cross-receiver evaluation and significantly improves robustness to receiver replacement.

Figures

Figures reproduced from arXiv: 2510.09405 by Fushuo Huo, Nan Cheng, Wenchao Xu, Xiucheng Wang, Yuhao Pan.

Figure 1
Figure 1. Figure 1: Illustration of transmitter and receiver hardware impairments. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the system model. are denoted by {De1 S , De2 S , . . . , Den S } and DeU , respectively. Through this separation, the source domain features exhibit increased domain invariance, thereby reducing the pairwise divergence between them, i.e., dH(Dei S , Dej S ) ≤ dH(Di S , D j S ), which in turn reduces ϵ. Moreover, by removing receiver￾specific features within each source domain, the target domai… view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity analysis of hyperparameters λ1, λ2, and λ3. indicating that excessive adversarial gradients may hinder the feature learning process and suppress discriminative transmit￾ter information. Next, we examine the influence of the regularization weight λ2 for the receiver-specific feature. As shown in [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

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

36 extracted references · 1 linked inside Pith

  1. [1]

    Internet of vehicles in big data era,

    W. Xu, H. Zhou, N. Cheng, F. Lyu, W. Shi, J. Chen, and X. Shen, “Internet of vehicles in big data era,”IEEE/CAA Journal of Automatica Sinica, vol. 5, no. 1, pp. 19–35, 2017

    2017

  2. [2]

    Connected vehicles: Solutions and challenges,

    N. Lu, N. Cheng, N. Zhang, X. Shen, and J. W. Mark, “Connected vehicles: Solutions and challenges,”IEEE internet of things journal, vol. 1, no. 4, pp. 289–299, 2014

    2014

  3. [3]

    Specific emitter identification based on hilbert–huang transform-based time–frequency–energy distri- bution features,

    Y . Yuan, Z. Huang, H. Wu, and X. Wang, “Specific emitter identification based on hilbert–huang transform-based time–frequency–energy distri- bution features,”IET communications, vol. 8, no. 13, pp. 2404–2412, 2014

    2014

  4. [4]

    Specific emitter identifica- tion via convolutional neural networks,

    L. Ding, S. Wang, F. Wang, and W. Zhang, “Specific emitter identifica- tion via convolutional neural networks,”IEEE communications letters, vol. 22, no. 12, pp. 2591–2594, 2018

    2018

  5. [5]

    Deep learning based rf fingerprint identification using differential constellation trace figure,

    L. Peng, J. Zhang, M. Liu, and A. Hu, “Deep learning based rf fingerprint identification using differential constellation trace figure,” IEEE Transactions on Vehicular Technology, vol. 69, no. 1, pp. 1091– 1095, 2019

    2019

  6. [6]

    Multimodal sparse representation-based classification scheme for rf fingerprinting,

    K. Yang, J. Kang, J. Jang, and H.-N. Lee, “Multimodal sparse representation-based classification scheme for rf fingerprinting,”IEEE Communications Letters, vol. 23, no. 5, pp. 867–870, 2019

    2019

  7. [7]

    Accurate and effi- cient wireless device fingerprinting using channel state information,

    J. Hua, H. Sun, Z. Shen, Z. Qian, and S. Zhong, “Accurate and effi- cient wireless device fingerprinting using channel state information,” in IEEE INFOCOM 2018-IEEE Conference on Computer Communications. IEEE, 2018, pp. 1700–1708

    2018

  8. [8]

    Radio fre- quency fingerprint identification for lora using spectrogram and cnn,

    G. Shen, J. Zhang, A. Marshall, L. Peng, and X. Wang, “Radio fre- quency fingerprint identification for lora using spectrogram and cnn,” in IEEE INFOCOM 2021-IEEE Conference on Computer Communications. IEEE, 2021, pp. 1–10

    2021

  9. [9]

    Robust wireless fin- gerprinting via complex-valued neural networks,

    S. Gopalakrishnan, M. Cekic, and U. Madhow, “Robust wireless fin- gerprinting via complex-valued neural networks,” in2019 IEEE Global Communications Conference (GLOBECOM). IEEE, 2019, pp. 1–6

    2019

  10. [10]

    Ex- posing the fingerprint: Dissecting the impact of the wireless channel on radio fingerprinting,

    A. Al-Shawabka, F. Restuccia, S. D’Oro, T. Jian, B. C. Rendon, N. Soltani, J. Dy, S. Ioannidis, K. Chowdhury, and T. Melodia, “Ex- posing the fingerprint: Dissecting the impact of the wireless channel on radio fingerprinting,” inIEEE INFOCOM 2020-IEEE Conference on Computer Communications. IEEE, 2020, pp. 646–655

    2020

  11. [11]

    Radiodiff: An effective generative diffusion model for sampling-free dynamic radio map construction,

    X. Wang, K. Tao, N. Cheng, Z. Yin, Z. Li, Y . Zhang, and X. Shen, “Radiodiff: An effective generative diffusion model for sampling-free dynamic radio map construction,”IEEE Transactions on Cognitive Communications and Networking, 2024

    2024

  12. [12]

    Radiodiff-k 2: Helmholtz equation informed generative diffu- sion model for multipath-aware radio map construction,

    X. Wang, Q. Zhang, N. Cheng, R. Sun, Z. Li, S. Cui, and X. Shen, “Radiodiff-k 2: Helmholtz equation informed generative diffu- sion model for multipath-aware radio map construction,”arXiv preprint arXiv:2504.15623, 2025

    arXiv 2025

  13. [13]

    Oracle: Optimized radio classification through convo- lutional neural networks,

    K. Sankhe, M. Belgiovine, F. Zhou, S. Riyaz, S. Ioannidis, and K. Chowdhury, “Oracle: Optimized radio classification through convo- lutional neural networks,” inIEEE INFOCOM 2019-IEEE conference on computer communications. IEEE, 2019, pp. 370–378

    2019

  14. [14]

    Towards scalable and channel-robust radio frequency fingerprint identification for lora,

    G. Shen, J. Zhang, A. Marshall, and J. R. Cavallaro, “Towards scalable and channel-robust radio frequency fingerprint identification for lora,” IEEE Transactions on Information Forensics and Security, vol. 17, pp. 774–787, 2022

    2022

  15. [15]

    Wisig: A large-scale wifi signal dataset for receiver and channel agnostic rf fingerprinting,

    S. Hanna, S. Karunaratne, and D. Cabric, “Wisig: A large-scale wifi signal dataset for receiver and channel agnostic rf fingerprinting,”IEEE Access, vol. 10, pp. 22 808–22 818, 2022

    2022

  16. [16]

    A deep learning approach to iot authentication,

    R. Das, A. Gadre, S. Zhang, S. Kumar, and J. M. Moura, “A deep learning approach to iot authentication,” in2018 IEEE international conference on communications (ICC). IEEE, 2018, pp. 1–6

    2018

  17. [17]

    Deep learning for rf device fingerprinting in cognitive communication networks,

    K. Merchant, S. Revay, G. Stantchev, and B. Nousain, “Deep learning for rf device fingerprinting in cognitive communication networks,”IEEE journal of selected topics in signal processing, vol. 12, no. 1, pp. 160– 167, 2018

    2018

  18. [18]

    Adaptive rf fingerprints fusion via dual attention convolutions,

    W. Zhang, W. Zhao, X. Tan, L. Shao, and C. Ran, “Adaptive rf fingerprints fusion via dual attention convolutions,”IEEE Internet of Things Journal, vol. 9, no. 24, pp. 25 181–25 195, 2022

    2022

  19. [19]

    Toward length-versatile and noise-robust radio frequency fingerprint identification,

    G. Shen, J. Zhang, A. Marshall, M. Valkama, and J. R. Cavallaro, “Toward length-versatile and noise-robust radio frequency fingerprint identification,”IEEE Transactions on Information Forensics and Secu- rity, vol. 18, pp. 2355–2367, 2023

    2023

  20. [20]

    Multi-channel attentive feature fusion for radio frequency fingerprinting,

    Y . Zeng, Y . Gong, J. Liu, S. Lin, Z. Han, R. Cao, K. Huang, and K. B. Letaief, “Multi-channel attentive feature fusion for radio frequency fingerprinting,”IEEE Transactions on Wireless Communications, vol. 23, no. 5, pp. 4243–4254, 2023

    2023

  21. [21]

    Equalization assisted domain adaptation for radio frequency fingerprint identification,

    R. Pan, H. Chen, H. Chen, and W.-Q. Wang, “Equalization assisted domain adaptation for radio frequency fingerprint identification,”IEEE Wireless Communications Letters, 2024

    2024

  22. [22]

    Channel-robust rf fingerprint identification using multi-task learning and receiver collabo- ration,

    Z. Chai, X. Peng, X. Huang, M. Li, and X. Yang, “Channel-robust rf fingerprint identification using multi-task learning and receiver collabo- ration,”IEEE Signal Processing Letters, 2024

    2024

  23. [23]

    Prototype- driven unsupervised domain adaptation for specific emitter identifica- tion,

    J. Chen, L. Yu, Y . Chen, X. Zheng, P. Chen, and K. Cheng, “Prototype- driven unsupervised domain adaptation for specific emitter identifica- tion,”IEEE Internet of Things Journal, 2024

    2024

  24. [24]

    Cross-domain, scalable, and in- terpretable rf device fingerprinting,

    T. Zhao, X. Wang, and S. Mao, “Cross-domain, scalable, and in- terpretable rf device fingerprinting,” inIEEE INFOCOM 2024-IEEE Conference on Computer Communications. IEEE, 2024, pp. 2099– 2108

    2024

  25. [25]

    Vc-sei: Robust variable-channel specific emitter identification method using semi-supervised domain adaptation,

    H. Wan, Q. Wang, X. Fu, Y . Wang, H. Zhao, Y . Lin, H. Sari, and G. Gui, “Vc-sei: Robust variable-channel specific emitter identification method using semi-supervised domain adaptation,”IEEE Transactions on Wireless Communications, 2024

    2024

  26. [26]

    Cross-receiver radio frequency fingerprint identification based on contrastive learning and subdomain adaptation,

    X. Zha, T. Li, Z. Qiu, and F. Li, “Cross-receiver radio frequency fingerprint identification based on contrastive learning and subdomain adaptation,”IEEE Signal Processing Letters, vol. 30, pp. 70–74, 2023

    2023

  27. [27]

    Cross-device radio frequency fingerprinting identification based on do- main adaptation,

    Z. Chen, Z. Pang, W. Hou, H. Wen, M. Wen, R. Zhao, and T. Tang, “Cross-device radio frequency fingerprinting identification based on do- main adaptation,”IEEE Transactions on Consumer Electronics, vol. 70, no. 1, pp. 2391–2400, 2024

    2024

  28. [28]

    Avoiding shortcuts: Enhancing channel-robust specific emitter identification via single-source domain generalization,

    Y . Wang, T. Ohtsuki, Z. Sun, D. Niyato, X. Wang, and G. Gui, “Avoiding shortcuts: Enhancing channel-robust specific emitter identification via single-source domain generalization,”IEEE Transactions on Wireless Communications, 2025

    2025

  29. [29]

    Gan-rxa: A practical scalable solution to receiver-agnostic transmitter fingerprinting,

    T. Zhao, S. Sarkar, E. Krijestorac, and D. Cabric, “Gan-rxa: A practical scalable solution to receiver-agnostic transmitter fingerprinting,”IEEE Transactions on Cognitive Communications and Networking, vol. 10, no. 2, pp. 403–416, 2023

    2023

  30. [30]

    Domain generaliza- tion for cross-receiver radio frequency fingerprint identification,

    Y . Zhang, Q. Li, H. Liu, L. Yang, and J. Yang, “Domain generaliza- tion for cross-receiver radio frequency fingerprint identification,”IEEE Internet of Things Journal, 2024

    2024

  31. [31]

    Detecting change in data streams,

    D. Kifer, S. Ben-David, and J. Gehrke, “Detecting change in data streams,” inVLDB, vol. 4. Toronto, Canada, 2004, pp. 180–191

    2004

  32. [32]

    Generalizing to unseen domains via distribution matching,

    I. Albuquerque, J. Monteiro, M. Darvishi, T. H. Falk, and I. Mitliagkas, “Generalizing to unseen domains via distribution matching,”arXiv preprint arXiv:1911.00804, 2019

    Pith/arXiv arXiv 1911

  33. [33]

    On learning invariant representations for domain adaptation,

    H. Zhao, R. T. Des Combes, K. Zhang, and G. Gordon, “On learning invariant representations for domain adaptation,” inInternational con- ference on machine learning. PMLR, 2019, pp. 7523–7532

    2019

  34. [34]

    A theory of learning from different domains,

    S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan, “A theory of learning from different domains,”Machine learning, vol. 79, pp. 151–175, 2010

    2010

  35. [35]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  36. [36]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. March, and V . Lempitsky, “Domain-adversarial training of neural networks,”Journal of machine learning research, vol. 17, no. 59, pp. 1–35, 2016

    2016

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.