Pith. sign in

REVIEW 3 major objections 5 minor 74 references

Shuffling for Semantic Secrecy

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Random shuffling of feature order can make an eavesdropper's channel capacity zero.

desk verdict A practical shuffling-based secrecy layer for semantic communication with strong experiments, but the zero-capacity proof does not cover the actual system. read the letter →

arxiv 2507.07401 v1 pith:JNEKJGEU submitted 2025-07-10 cs.CR

classification cs.CR
keywords semanticcommunicationwiretapchannelrandomshufflingsharedsecretkeyleakageratesecrecycapacitypermutationequivariancephysicallayersecurity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the order in which semantic features are transmitted can itself serve as the secret key, and that randomly shuffling those features before transmission drives the eavesdropper's achievable rate to zero while the legitimate receiver, knowing the permutation, recovers the message normally. The target is a wiretap channel in a learned semantic communication system: Alice encodes text, image, or speech into feature tensors, permutes rows and/or columns with a shared random key, and transmits; Bob inverts the permutation and decodes. The paper claims this gives near-zero information leakage to Eve with negligible extra computation, and experimentally shows that Bob's reconstruction quality stays high under both AWGN and Rayleigh channels while Eve's recovered sentences, images, or speech contain no useful semantic content. The practical payoff is a secrecy layer that plugs into existing semantic encoders rather than requiring a learned encryption network or a long one-time pad.

What carries the argument

The load-bearing object is the noisy permutation channel: an $n$-symbol input is randomly permuted at a chosen grain $g$ before passing through a discrete memoryless channel, so an observer without the key sees the same multiset of symbols in a random order. The paper adapts this to semantic features by making the permutation pattern the shared key $(K_R, K_C)$, where row shuffling permutes the $N$ token or patch positions, column shuffling permutes the $L$ feature dimensions, and audio features admit a third filter-level shuffle contributing $C!$ additional choices. Transformer and MLP permutation equivariance guarantees that the legitimate decoder can undo the shuffle exactly at negligible extra FLOPs, while the same property restricts CNN-based audio systems to shuffling at the channel-encoder output, where convolution is not permutation-equivariant.

What would settle it

Train an eavesdropper on the actual continuous-valued shuffled features from the authors' AWGN or Rayleigh setup, giving it no permutation key but full knowledge of the feature distribution, and measure whether its reconstruction BLEU, PSNR, or SDR stays at the random-guess level as blocklength grows; any systematic recovery above that baseline would contradict the claimed $C_{\mathrm{Eve}}=0$ for the implemented channels.

Watch

Extended reading notes

Core claim

The central claim is that, in the limit of long blocklengths, the eavesdropper's channel capacity is exactly zero: when the shuffle pattern is unknown, the noisy permutation channel linking Alice's encoded message to Eve's received signal has transmission rate $R \leq 0$, so $C_{\mathrm{Eve}} = 0$. The proof, given for a discrete memoryless channel, treats the shuffle as a discrete channel that maps each $n$-length input to a uniformly chosen permutation of its $g$-sized grains; the number of possible shuffles grows factorially in $n/g$, so the per-symbol mutual information contributed by ordering vanishes as $n \to \infty$. Secrecy therefore comes from the combinatorial size of the key space, $|K| = N!L!$ for matrix features and $N!L!C!$ for audio tensors, rather than from channel noise or from a trained encryption network. Because the shuffle is exactly reversible, Bob's channel capacity is unaffected, and the paper illustrates this with a binary symmetric channel where $C_{\mathrm{Bob}} = 1-H(p)$ while $C_{\mathrm{Eve}} = 0$.

Load-bearing premise

The zero-capacity proof assumes the transmission channel is a discrete memoryless channel with a finite alphabet, but the implemented system transmits continuous-valued feature vectors over AWGN and Rayleigh channels; if an eavesdropper can exploit statistical structure in the continuous multiset of feature values, the claimed zero leakage may not hold.

Editorial extensions

If this is right

  • Bob's channel capacity is preserved exactly because the shuffle is reversible, so the scheme can add secrecy without sacrificing the legitimate rate.
  • The overhead of shuffling is small: the paper reports added FLOPs of 0.04% for text, 0.36% for speech, and 3.51% for image transmission.
  • On text, image, and speech tasks, the method keeps Bob's BLEU, PSNR, and SDR high while Eve's reconstruction quality stays near random-guess level under both AWGN and Rayleigh fading channels.
  • The key rate can be tuned through shuffle grain and the choice of row, column, or filter shuffling, with higher key rates yielding lower Eve reconstruction quality across all three modalities.
  • Because it relies on permutation equivariance rather than a learned encryption module, the shuffle layer can be attached to existing semantic communication encoders as a plugin.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The zero-capacity theorem is proven for a discrete memoryless channel, but the implemented system sends continuous-valued features over AWGN and Rayleigh channels; whether the multiset of continuous feature values leaks information is an open question that the paper's experiments do not fully settle.
  • A natural testable extension is an eavesdropper that performs histogram or multiset matching on the received continuous features instead of end-to-end neural decoding; if such an attacker recovers above random-guess quality, a symbol-level randomization or quantization layer would be needed to restore the guarantee.
  • Since the key length grows as $\log(N!L!)$, the scheme behaves like a one-time pad for the ordering but not for the symbol values, suggesting that combining shuffling with per-symbol noise could approach information-theoretic secrecy over continuous channels.
  • The same permutation-equivariance route could extend to other order-sensitive modalities such as video frames or point clouds, or be combined with physical-layer security to protect against attackers that know the feature distribution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a secure semantic communication system in which the random permutation pattern of DNN intermediate features serves as a shared secret key between Alice and Bob. The encoder/decoder pair is trained end-to-end with a loss that combines reconstruction quality, a MINE-based leakage term, and a secrecy-capacity term, while Eve is modeled as a decoder trained by inversion attack without the key. The central theoretical claim is that Eve's channel capacity is zero, based on an analogy to noisy permutation channels (Section V, Theorem 2). Experiments on text, image, and speech transmission over AWGN and Rayleigh channels show that Bob's reconstruction quality is high while Eve's reconstruction metrics are low, together with ablations of shuffle position, grain, and key rate.

Significance. If the zero-capacity claim applied to the implemented system, the paper would make a substantial contribution: a training-free, low-overhead secrecy mechanism that plugs into existing semantic communication systems and avoids the black-box reliability issues of neural encryption. The paper's strengths are its extensive multimodal experiments, sensible baselines, and the correct identification of the noisy-permutation-channel connection, whose capacity-zero result is explicitly attributed to prior work. However, the theoretical guarantee is proven only for a finite-alphabet DMC, while the evaluated system transmits continuous-valued features; the main security claim is therefore not established for the actual channels. The practical evidence of low leakage is credible but is empirical and should not be presented as a consequence of Theorem 2.

major comments (3)
  1. [Section V, Eq. (14); Supplementary A, Eq. (17)] Theorem 2 is proved only under the finite-alphabet DMC assumption stated just before it. In the evaluated system of Section VI, X is a continuous-valued feature matrix (e.g., X in R^{N x V} for text/image and X in R^{C x N x L} for speech) transmitted over AWGN and Rayleigh channels. For a continuous alphabet, log|X| in the bound is infinite and the transition-matrix argument in Supplementary Eq. (16), which counts finitely many output sequences, does not apply. Consequently the statement 'we have CEve = 0' cannot be inferred for the actual system; the unordered multiset of continuous feature values can carry positive information, for instance through the empirical distribution of the received values. Please either restrict the zero-capacity claim to a quantized finite-alphabet version and state explicitly that the experimental system lies outside the theorem, or provide a continuous-alphabet analysis and weaken the claim accordingly.
  2. [Supplementary A, Eq. (16)] The transition matrix P_{X^n|U^n} is written with each row containing exactly a=(n/g)! nonzero entries of probability 1/a. This is not true when the input vector contains repeated symbols: the number of distinct outputs under permutation is smaller and the probabilities are non-uniform. The final bound can likely be repaired by replacing the exact entropy computation with H(X^n|U^n) <= log a, but the proof as written is incorrect. In addition, Eq. (15) jumps from I(M; M_hat) to I(U^n; X^n); the intermediate step I(M; M_hat) <= I(U^n; Z^n) <= I(U^n; X^n) via data processing should be stated explicitly.
  3. [Section IV-A, Eq. (8), and Fig. 8] The MINE estimator is a lower bound on mutual information, as the paper itself notes. Reporting this estimate as the 'information leakage rate RL' therefore does not provide an upper bound and cannot by itself certify near-zero leakage. The direct Eve-reconstruction metrics (BLEU, PSNR, SDR, sentence similarity) are the more convincing evidence of low leakage and should be the primary support; if RL is reported, it should be labeled as a lower-bound estimate or complemented by an upper-bound estimator.
minor comments (5)
  1. [Section V, Theorem 2] The sentence 'letting n to infinity produces R <= 0' should read 'R_Eve <= 0'; the symbol R is used earlier for the transmission rate R = log|M|/n, and the two should not be conflated.
  2. [Section III-A] The sentence 'Specifically, we assume that the input of the communication system is text' is immediately followed by a description covering images and audio; this sentence should be generalized or moved to the text-specific instantiation.
  3. [Supplementary A, Eq. (17)] The text says step (a) follows from the uniform-distribution assumption of X^n, but the derivation assumed a uniform distribution on U^n; clarify which random variable is uniform and state explicitly that the proof is discrete-only (or handle the continuous case separately).
  4. [Section VI-G and Fig. 14] The observed exceptions to the monotonic relation between key rate and secrecy (row vs column, 1D g=8 vs 1D g=12) are explained by heuristic arguments about token semantics and channel-encoder dimension; these explanations are plausible but post hoc, and the text should avoid presenting the monotonic relation as a theoretical prediction.
  5. [Table II] The term 'Floating Point Operations per Second' is a misnomer for FLOPs; the table reports per-inference operation counts, not a rate, and should be worded accordingly.

Circularity Check

0 steps flagged · score 2.0 of 10

The central CEve=0 result rests on independent noisy-permutation-channel theory, not on fitted parameters; the only self-citation is minor and not load-bearing.

full rationale

The claimed derivation of CEve=0 (Theorem 2, Section V, with proof in Supplementary A) is not circular. It starts from the Markov chain U^n-X^n-Z^n, applies Fano's inequality and data processing, and bounds I(U^n;X^n) by directly inspecting the uniform-permutation transition matrix in Eqs. (15)-(17). No parameter is fitted to the data whose secrecy is being claimed, and the zero-capacity conclusion is the same phenomenon previously established for noisy permutation channels by independent authors [9], [54]. The self-citation to [55] for Transformer permutation equivariance is not load-bearing for the security theorem: it only justifies the flexibility of placing the row shuffle at positions 1-3, and the equivariance property is externally checkable rather than assumed into the central result. The only mild self-evaluation element is the MINE-estimated leakage rate in Fig. 8, since the same RL term is minimized in the training loss of Eq. (11) and estimated by a network trained jointly with the backbone; however, the paper also reports independent leakage evidence (Eve's BLEU, PSNR, SDR, and visualizations), so the conclusion does not reduce to the fitted estimator. The finite-alphabet DMC restriction stated in Section V is a real scope gap relative to the continuous AWGN/Rayleigh evaluation, but that is a correctness or transfer concern, not circularity. Overall, the central derivation is self-contained against external benchmarks, and no circular reduction was found.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central security claim rests on two external results (noisy permutation channel capacity, Transformer equivariance), one ad hoc channel assumption that does not match the experiments, and a threat model plus an MI estimator whose reliability is assumed. No new physical entities are introduced.

free parameters (5)
  • alpha (leakage weight) = 0.01
    Weight factor for the leakage rate term in the loss (Eq. 11); varied in ablation to 0, 0.01, 0.1, 0.5, 0.8.
  • beta (secrecy capacity weight) = 0.01
    Weight factor for the secrecy capacity term in the loss (Eq. 11); varied in ablation to 0, 0.001, 0.01, 0.1, 0.2.
  • gamma (row shuffling position) = 1 (Transformer), 3 (CNN)
    Position at which row shuffling is applied; the paper argues the choice has minor impact due to permutation equivariance and explores positions 1, 2, 3.
  • g (shuffle grain) = 1
    Smallest unit of permutation; varied in text experiments to 2, 4, 8, 12, 16.
  • training noise variance = 0.1
    Gaussian noise variance used when training the encoder-decoder (Section VI-A); fixed during training, while test SNR varies.
assumptions (5)
  • standard math Transformer encoder row permutation equivariance: Enc(PR U) = PR Enc(U) (Eq. 4a)
    Used to justify applying row shuffling at any of three positions and reversing it at the receiver. Cited from [55]; not proved in this paper.
  • standard math Noisy permutation channel capacity is zero for a finite-alphabet DMC (from [9], [54])
    Theorem 2 applies this known result to bound Eve's rate. The paper's proof in Supplementary A is a simplified restatement.
  • ad hoc to paper The transmission channel is a discrete memoryless channel with finite alphabet for the leakage analysis
    Explicitly assumed in Section V for ease of analysis, but contradicted by the continuous AWGN and Rayleigh channel experiments.
  • domain assumption Eve is limited to the same decoder architecture as Bob and lacks only the key
    Defines the threat model in Section III-A; excludes stronger attackers that might exploit known plaintext or different architectures.
  • domain assumption MINE provides a valid estimate of the mutual information terms
    The loss and the reported leakage rate rely on the Donsker-Varadhan lower bound; the paper does not quantify estimator bias or variance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Shuffling for Semantic Secrecy." pith.science (2026). https://pith.science/paper/JNEKJGEU

@misc{pith2026250707401,
  author       = {Pith},
  title        = {Pith review of: Shuffling for Semantic Secrecy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JNEKJGEU}},
  note         = {Machine review of arXiv:2507.07401}
}
read the original abstract

Deep learning draws heavily on the latest progress in semantic communications. The present paper aims to examine the security aspect of this cutting-edge technique from a novel shuffling perspective. Our goal is to improve upon the conventional secure coding scheme to strike a desirable tradeoff between transmission rate and leakage rate. To be more specific, for a wiretap channel, we seek to maximize the transmission rate while minimizing the semantic error probability under the given leakage rate constraint. Toward this end, we devise a novel semantic security communication system wherein the random shuffling pattern plays the role of the shared secret key. Intuitively, the permutation of feature sequences via shuffling would distort the semantic essence of the target data to a sufficient extent so that eavesdroppers cannot access it anymore. The proposed random shuffling method also exhibits its flexibility in working for the existing semantic communication system as a plugin. Simulations demonstrate the significant advantage of the proposed method over the benchmark in boosting secure transmission, especially when channels are prone to strong noise and unpredictable fading.

Figures

Figures reproduced from arXiv: 2507.07401 by the authors.

Figure 1
Figure 1. Traditional communication system versus semantic communication [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Secure communication over a public channel with a shared key. The symbol [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. (a) The detailed illustration of permutation and inverse permutation. (b) An example of a single row permutation and column permutation cannot obtain an arbitrary arrangement pattern. data of different modalities. Hence we instantiate our frame￾work with three types of data, i.e., text, image, and audio stream, respectively. For different types of data, the specific meaning of each notation and the model structure i… view at source ↗
Figures from the paper (17 more)
Figure 4
Figure 4. Figure 4: illustration of the communication system with shuffle operation [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Eve’s inversion attack results without protection for different types of data. In the following sections, we show Eve’s reconstruction results by data types under our approach and other baselines. D. Text Reconstruction Results Fig. 7a and 7b illustrate the BLEU score …
Figure 6
Figure 6. Figure 6: CelebA-HQ Dataset: Visualization of Eve’s inversion attack results without any protection. and Rayleigh channels. Note that the traditional method uses BPSK as the digital modulation scheme when the channel SNR is below 6dB according to the MCS table specifications, wh…
Figure 7
Figure 7. Figure 7: Bob and Eve’s BLEU scores for our approach compared with the traditional method, adversarial learning-based method, and LWE method in (a) AWGN channel, and (b) Rayleigh fading channel. 0 3 6 9 12 15 18 SNR(dB) 0.00 0.01 0.02 0.03 Information Leakage Rate our method adv…
Figure 8
Figure 8. Figure 8: Information Leakage Rates RL of different methods under AWGN channel. the optimal performance and DeepSC-S has a close perfor￾mance to ours, affirming that the shuffling operations do not compromise the reconstruction accuracy of speech data. Compared to the semantic a…
Figure 11
Figure 11. Figure 11: Image transmission on CIFAR-100: PSNR scores of Eve in different shuffling modes. Origin Eve in col shuffle Eve in row shuffle Eve in row-col shuffle [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: demonstrates that, under pure column shuffling, Eve can approximately reconstruct the color distribution of the image. This is because column shuffling destroys the spatial relationship within each ViT patch but retains the relationship between patches. With row shuff…
Figure 13
Figure 13. Figure 13: Speech transmission: Eve’s spectrogram differential heat map in [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: Text transmission: Eve’s BLEU score (1-grams) with different key [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]
Figure 16
Figure 16. Figure 16: Bob’s and Eve’s BLEUs at β=0.01 and varying αs. 0 3 6 9 12 15 18 SNR (dB) 0.0 0.2 0.4 0.6 0.8 1.0 BLEU (2-grams) =0 =0.001 =0.01 =0.1 =0.2 (a) Bob. 0 3 6 9 12 15 18 SNR (dB) 0.025 0.030 0.035 0.040 0.045 (b) Eve [PITH_FULL_IMAGE:figures/full_fig_p014_16.png]
Figure 17
Figure 17. Figure 17: Bob’s and Eve’s BLEUs at α=0.01 and varying βs. The results of fixing α = 0.01 and varying the value of β are shown in [PITH_FULL_IMAGE:figures/full_fig_p014_17.png]
Figure 20
Figure 20. Figure 20: displays the MS-SSIM score on CelebA-HQ versus SNRs across different channels, where our method consistently 0 2 4 6 8 10 12 14 16 18 20 SNR (dB) 0.2 0.4 0.6 0.8 1.0 MS-SSIM Our: Bob Our: eve WITT Trad: Bob Trad: Eve LWE: Bob LWE: Eve (a) AWGN channel. 0 2 4 6 8 10 12…
Figure 21
Figure 21. Figure 21: presents the inversion attack results on unprotected CIFAR-100, following a similar setting to that of [PITH_FULL_IMAGE:figures/full_fig_p018_21.png]
Figure 22
Figure 22. Figure 22: Image transmission on CelebA-HQ: PSNR scores of Eve in different shuffling modes [PITH_FULL_IMAGE:figures/full_fig_p018_22.png]
Figure 25
Figure 25. Figure 25: Image transmission on CIFAR-100: Bob and Eve’s MS-SSIM Scores for different methods in (a) AWGN channel and (b) Rayleigh channel. of the traditional method under the AWGN channel is close to that of the semantic approach due to the low resolution of CIFAR-100 images. …
Figure 26
Figure 26. Figure 26: Speech transmission: Bob and Eve’s PESQ score for different methods [PITH_FULL_IMAGE:figures/full_fig_p019_26.png]
Figure 24
Figure 24. Figure 24: Image transmission on CIFAR-100: Bob and Eve’s PSNR scores for different methods in (a) AWGN channel and (b) Rayleigh channel [PITH_FULL_IMAGE:figures/full_fig_p019_24.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 65 canonical work pages

  1. [1]

    Deep learning enabled semantic-secure communication with shuffling,

    F. Chen, L. Xiang, H. V . Cheng, and K. Shen, “Deep learning enabled semantic-secure communication with shuffling,” in Proc. IEEE Global Commun. Conf. (GLOBECOM) , 2023, pp. 6338–6343

  2. [2]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,” Bell Syst. Tech. J., vol. 27, no. 3, pp. 379–423, July 1948

  3. [3]

    Semantic communications: Principles and challenges,

    Z. Qin, X. Tao, J. Lu, W. Tong, and G. Y . Li, “Semantic communications: Principles and challenges,” arXiv:2201.01389, 2021

  4. [4]

    A method for obtaining digital signatures and public-key cryptosystems,

    R. L. Rivest, A. Shamir, and L. Adleman, “A method for obtaining digital signatures and public-key cryptosystems,”Commun. ACM, vol. 21, no. 2, pp. 120–126, 1978

  5. [5]

    Wire- less information-theoretic security,

    M. Bloch, J. Barros, M. R. D. Rodrigues, and S. W. McLaughlin, “Wire- less information-theoretic security,” IEEE Trans. Inf. Theory , vol. 54, no. 6, pp. 2515–2534, 2008

  6. [6]

    Encrypted semantic communi- cation using adversarial training for privacy preserving,

    X. Luo, Z. Chen, M. Tao, and F. Yang, “Encrypted semantic communi- cation using adversarial training for privacy preserving,” IEEE Commun. Lett., vol. 27, no. 6, pp. 1486–1490, June 2023

  7. [7]

    Semprotector: A unified framework for semantic protection in deep learning-based semantic communication systems,

    X. Liu, G. Nan, Q. Cui, Z. Li, P. Liu, Z. Xing, H. Mu, X. Tao, and T. Q. S. Quek, “Semprotector: A unified framework for semantic protection in deep learning-based semantic communication systems,” IEEE Commun. Mag., vol. 61, no. 11, pp. 56–62, November 2023

  8. [8]

    Privacy-preserving task-oriented semantic communications against model inversion attacks,

    Y . Wang, S. Guo, Y . Deng, H. Zhang, and Y . Fang, “Privacy-preserving task-oriented semantic communications against model inversion attacks,” IEEE Trans. Wireless Commun., vol. 23, no. 8, pp. 10 150–10 165, 2024

Show all 74 references
  1. [9]

    Capacity of noisy permutation channels,

    J. Tang and Y . Polyanskiy, “Capacity of noisy permutation channels,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT) , June 2022, pp. 1987–1992

  2. [10]

    Communication theory of secrecy systems,

    C. E. Shannon, “Communication theory of secrecy systems,” Bell Syst. Tech. J., vol. 28, no. 4, pp. 656–715, Oct 1949

  3. [11]

    The wire-tap channel,

    A. D. Wyner, “The wire-tap channel,” Bell Syst. Tech. J., vol. 54, no. 8, pp. 1355–1387, Oct 1975

  4. [12]

    Wiretap channel with secure rate-limited feedback,

    E. Ardestanizadeh, M. Franceschetti, T. Javidi, and Y .-H. Kim, “Wiretap channel with secure rate-limited feedback,” IEEE Trans. Inf. Theory , vol. 55, no. 12, pp. 5353–5361, Dec 2009

  5. [13]

    Polar coding for the wiretap channel with shared key,

    H. Wang, X. Tao, N. Li, and Z. Han, “Polar coding for the wiretap channel with shared key,” IEEE Trans. Inf. Forensics Security , vol. 13, no. 6, pp. 1351–1360, June 2018

  6. [14]

    An ldpc code based physical layer message authentication scheme with prefect security,

    D. Chen, N. Zhang, R. Lu, X. Fang, K. Zhang, Z. Qin, and X. Shen, “An ldpc code based physical layer message authentication scheme with prefect security,” IEEE J. Sel. Areas Commun. , vol. 36, no. 4, pp. 748– 761, April 2018

  7. [15]

    Generating a binary symmetric channel for wiretap codes,

    W. K. Harrison, T. Fernandes, M. A. C. Gomes, and J. P. Vilela, “Generating a binary symmetric channel for wiretap codes,” IEEE Trans. Inf. Forensics Security, vol. 14, no. 8, pp. 2128–2138, Aug 2019

  8. [16]

    Joint precoding and artificial noise design for mu-mimo wiretap channels,

    E. Choi, M. Oh, J. Choi, J. Park, N. Lee, and N. Al-Dhahir, “Joint precoding and artificial noise design for mu-mimo wiretap channels,” IEEE Trans. Commun. , vol. 71, no. 3, pp. 1564–1578, March 2023

  9. [17]

    Deep learning enabled semantic communication systems,

    H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Trans. Signal Process., vol. 69, pp. 2663–2675, 2021

  10. [18]

    Deep learning for joint source- channel coding of text,

    N. Farsad, M. Rao, and A. Goldsmith, “Deep learning for joint source- channel coding of text,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), April 2018, pp. 2326–2330

  11. [19]

    A lite distributed semantic communication system for internet of things,

    H. Xie and Z. Qin, “A lite distributed semantic communication system for internet of things,” IEEE J. Sel. Areas Commun. , vol. 39, no. 1, pp. 142–153, Jan 2021

  12. [20]

    An investigation on intelligent relay assisted semantic communication networks,

    S. Ma, W. Liang, B. Zhang, and D. Wang, “An investigation on intelligent relay assisted semantic communication networks,” in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC) , March 2023, pp. 1–6

  13. [21]

    Deep source-channel coding for sentence semantic transmission with harq,

    P. Jiang, C.-K. Wen, S. Jin, and G. Y . Li, “Deep source-channel coding for sentence semantic transmission with harq,” IEEE Trans. Commun. , vol. 70, no. 8, pp. 5225–5240, Aug 2022

  14. [22]

    Toward semantic communications: Deep learning-based image semantic coding,

    D. Huang, F. Gao, X. Tao, Q. Du, and J. Lu, “Toward semantic communications: Deep learning-based image semantic coding,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 55–71, Jan 2023

  15. [23]

    Predictive and adaptive deep coding for wireless image transmission in semantic communication,

    W. Zhang, H. Zhang, H. Ma, H. Shao, N. Wang, and V . C. M. Leung, “Predictive and adaptive deep coding for wireless image transmission in semantic communication,” IEEE Trans. Wireless Commun. , vol. 22, no. 8, pp. 5486–5501, Aug 2023. 15

  16. [24]

    Semantic communication system based on semantic slice models propagation,

    C. Dong, H. Liang, X. Xu, S. Han, B. Wang, and P. Zhang, “Semantic communication system based on semantic slice models propagation,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 202–213, Jan 2023

  17. [25]

    Deep learning- enabled semantic communication systems with task-unaware transmitter and dynamic data,

    H. Zhang, S. Shao, M. Tao, X. Bi, and K. B. Letaief, “Deep learning- enabled semantic communication systems with task-unaware transmitter and dynamic data,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 170–185, Jan 2023

  18. [26]

    On the role of vit and cnn in semantic communications: Analysis and prototype validation,

    H. Yoo, L. Dai, S. Kim, and C.-B. Chae, “On the role of vit and cnn in semantic communications: Analysis and prototype validation,” IEEE Access, vol. 11, pp. 71 528–71 541, 2023

  19. [27]

    Witt: A wireless image transmission transformer for semantic communications,

    K. Yang, S. Wang, J. Dai, K. Tan, K. Niu, and P. Zhang, “Witt: A wireless image transmission transformer for semantic communications,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP) , June 2023, pp. 1–5

  20. [28]

    Semantic communication systems for speech transmission,

    Z. Weng and Z. Qin, “Semantic communication systems for speech transmission,” IEEE J. Sel. Areas Commun. , vol. 39, no. 8, pp. 2434– 2444, Aug 2021

  21. [29]

    Semantic-preserved communication system for highly efficient speech transmission,

    T. Han, Q. Yang, Z. Shi, S. He, and Z. Zhang, “Semantic-preserved communication system for highly efficient speech transmission,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 245–259, Jan 2023

  22. [30]

    Deep learning enabled semantic communications with speech recognition and synthesis,

    Z. Weng, Z. Qin, X. Tao, C. Pan, G. Liu, and G. Y . Li, “Deep learning enabled semantic communications with speech recognition and synthesis,” IEEE Trans. Wireless Commun. , vol. 22, no. 9, pp. 6227– 6240, Sep. 2023

  23. [31]

    Wireless deep video semantic transmission,

    S. Wang, J. Dai, Z. Liang, K. Niu, Z. Si, C. Dong, X. Qin, and P. Zhang, “Wireless deep video semantic transmission,” IEEE J. Sel. Areas Commun., vol. 41, no. 1, pp. 214–229, Jan 2023

  24. [32]

    Toward adaptive semantic communications: Efficient data transmission via online learned nonlinear transform source-channel coding,

    J. Dai, S. Wang, K. Yang, K. Tan, X. Qin, Z. Si, K. Niu, and P. Zhang, “Toward adaptive semantic communications: Efficient data transmission via online learned nonlinear transform source-channel coding,” IEEE J. Sel. Areas Commun. , vol. 41, no. 8, pp. 2609–2627, Aug 2023

  25. [33]

    Wireless semantic communi- cations for video conferencing,

    P. Jiang, C.-K. Wen, S. Jin, and G. Y . Li, “Wireless semantic communi- cations for video conferencing,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 230–244, Jan 2023

  26. [34]

    Model division multiple access for semantic communications,

    P. Zhang, X. Xu, C. Dong, K. Niu, H. Liang, Z. Liang, X. Qin, M. Sun, H. Chen, N. Ma, W. Xu, G. Wang, and X. Tao, “Model division multiple access for semantic communications,” Frontiers of Information Technology & Electronic Engineering, vol. 24, no. 6, pp. 801–812, Jun. 2023

  27. [35]

    Deepma: End-to-end deep multiple access for wireless image transmission in semantic communication,

    W. Zhang, K. Bai, S. Zeadally, H. Zhang, H. Shao, H. Ma, and V . C. M. Leung, “Deepma: End-to-end deep multiple access for wireless image transmission in semantic communication,” IEEE Trans. on Cogn. Commun. Netw., vol. 10, no. 2, pp. 387–402, 2024

  28. [36]

    Orthogonal model division multiple access,

    H. Liang, K. Liu, X. Liu, H. Jiang, C. Dong, X. Xu, K. Niu, and P. Zhang, “Orthogonal model division multiple access,” IEEE Trans. Wireless Commun., vol. 23, no. 9, pp. 11 693–11 707, 2024

  29. [37]

    Rethinking wireless communication security in semantic internet of things,

    H. Du, J. Wang, D. Niyato, J. Kang, Z. Xiong, M. Guizani, and D. I. Kim, “Rethinking wireless communication security in semantic internet of things,” IEEE Wireless Commun., vol. 30, no. 3, pp. 36–43, 2023

  30. [38]

    Secure semantic communications: Challenges, approaches, and opportunities,

    M. Shen, J. Wang, H. Du, D. Niyato, X. Tang, J. Kang, Y . Ding, and L. Zhu, “Secure semantic communications: Challenges, approaches, and opportunities,” IEEE Netw., vol. 38, no. 4, pp. 197–206, 2024

  31. [39]

    Secure semantic communications: Fundamentals and challenges,

    Z. Yang, M. Chen, G. Li, Y . Yang, and Z. Zhang, “Secure semantic communications: Fundamentals and challenges,” 2023. [Online]. Available: https://arxiv.org/abs/2301.01421

  32. [40]

    Physical-layer adversarial robustness for deep learning-based semantic communications,

    G. Nan, Z. Li, J. Zhai, Q. Cui, G. Chen, X. Du, X. Zhang, X. Tao, Z. Han, and T. Q. S. Quek, “Physical-layer adversarial robustness for deep learning-based semantic communications,” IEEE J. Sel. Areas Commun., vol. 41, no. 8, pp. 2592–2608, 2023

  33. [41]

    Secure semantic communication via paired adversarial residual networks,

    B. He, F. Wang, and T. Q. S. Quek, “Secure semantic communication via paired adversarial residual networks,” IEEE Wireless Commun. Lett., vol. 13, no. 10, pp. 2832–2836, 2024

  34. [42]

    Deep joint source- channel coding for image transmission with visual protection,

    J. Xu, B. Ai, W. Chen, N. Wang, and M. Rodrigues, “Deep joint source- channel coding for image transmission with visual protection,” IEEE Trans. on Cogn. Commun. Netw. , vol. 9, no. 6, pp. 1399–1411, 2023

  35. [43]

    Secure transmission in wireless semantic communications with adversarial training,

    J. Shi, Q. Zhang, W. Zeng, S. Li, and Z. Qin, “Secure transmission in wireless semantic communications with adversarial training,” IEEE Commun. Lett., pp. 1–1, 2025

  36. [44]

    Wireless image transmission with semantic and security awareness,

    M. Zhang, Y . Li, Z. Zhang, G. Zhu, and C. Zhong, “Wireless image transmission with semantic and security awareness,” 2022. [Online]. Available: https://arxiv.org/abs/2212.00227

  37. [45]

    Secure semantic com- munication for image transmission in the presence of eavesdroppers,

    S. Tang, C. Liu, Q. Yang, S. He, and D. Niyato, “Secure semantic com- munication for image transmission in the presence of eavesdroppers,” in Proc. IEEE Global Commun. Conf. (GLOBECOM) , 2024, pp. 2172– 2177

  38. [46]

    Deep joint source-channel and encryption coding: Secure semantic communications,

    T.-Y . Tung and D. Gündüz, “Deep joint source-channel and encryption coding: Secure semantic communications,” in Proc. IEEE Int. Commun. Conf. (ICC), May 2023, pp. 5620–5625

  39. [47]

    Secur- ing semantic communications with physical-layer semantic encryption and obfuscation,

    Q. Qin, Y . Rong, G. Nan, S. Wu, X. Zhang, Q. Cui, and X. Tao, “Secur- ing semantic communications with physical-layer semantic encryption and obfuscation,” in Proc. IEEE Int. Commun. Conf. (ICC) , May 2023, pp. 5608–5613

  40. [48]

    A superposition code-based semantic communication approach with quantifiable and controllable security,

    W. Chen, S. Shao, Q. Yang, Z. Zhang, and P. Zhang, “A superposition code-based semantic communication approach with quantifiable and controllable security,” 2025. [Online]. Available: https://arxiv.org/abs/2401.13980

  41. [49]

    Semantic communication-assisted physical layer security over fading wiretap channels,

    X. Mu and Y . Liu, “Semantic communication-assisted physical layer security over fading wiretap channels,” in Proc. IEEE Int. Commun. Conf. (ICC), 2024, pp. 2101–2106

  42. [50]

    Semantic entropy can simultaneously benefit transmission efficiency and channel security of wireless semantic communications,

    Y . Rong, G. Nan, M. Zhang, S. Chen, S. Wang, X. Zhang, N. Ma, S. Gong, Z. Yang, Q. Cui, X. Tao, and T. Q. S. Quek, “Semantic entropy can simultaneously benefit transmission efficiency and channel security of wireless semantic communications,” IEEE Trans. Inf. Forensics Securi...

  43. [51]

    The model inversion eavesdrop- ping attack in semantic communication systems,

    Y . Chen, Q. Yang, Z. Shi, and J. Chen, “The model inversion eavesdrop- ping attack in semantic communication systems,” in Proc. IEEE Global Commun. Conf. (GLOBECOM) , 2023, pp. 5171–5177

  44. [52]

    Neural network inversion in adversarial setting via background knowledge alignment,

    Z. Yang, J. Zhang, E.-C. Chang, and Z. Liang, “Neural network inversion in adversarial setting via background knowledge alignment,” in Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, 2019, pp. 225–240

  45. [53]

    Model inversion attacks that exploit confidence information and basic countermeasures,

    M. Fredrikson, S. Jha, and T. Ristenpart, “Model inversion attacks that exploit confidence information and basic countermeasures,” in Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, 2015, pp. 1322–1333

  46. [54]

    Coding theorems for noisy permutation channels,

    A. Makur, “Coding theorems for noisy permutation channels,” IEEE Trans. Inf. Theory, vol. 66, no. 11, pp. 6723–6748, Nov 2020

  47. [55]

    Permutation equiv- ariance of transformers and its applications,

    H. Xu, L. Xiang, H. Ye, D. Yao, P. Chu, and B. Li, “Permutation equiv- ariance of transformers and its applications,” in IEEE/CVF Computer Vision and Pattern Recognition Conference (CVPR ’24) , 2024

  48. [56]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv:2010.11929, 2020

  49. [57]

    El Gamal and Y .-H

    A. El Gamal and Y .-H. Kim, Network information theory . Cambridge university press, 2011

  50. [58]

    Mutual information neural estimation,

    M. I. Belghazi, A. Baratin, S. Rajeshwar, S. Ozair, Y . Bengio, A. Courville, and D. Hjelm, “Mutual information neural estimation,” in Proc.35th Int. Conf. Mach. Learn. , vol. 80, 10–15 Jul 2018, pp. 531– 540

  51. [59]

    Europarl: A parallel corpus for statistical machine translation,

    P. Koehn, “Europarl: A parallel corpus for statistical machine translation,” in Proc. AAMT Mach. Transl. Summit. , Phuket, Thailand, Sep. 13-15 2005, pp. 79–86. [Online]. Available: https://aclanthology. org/2005.mtsummit-papers.11

  52. [60]

    Progressive growing of GANs for improved quality, stability, and variation,

    T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive growing of GANs for improved quality, stability, and variation,” in International Conference on Learning Representations , 2018

  53. [61]

    Noisy reverberant speech database for training speech enhancement algorithms and tts models,

    C. Valentini-Botinhao, “Noisy reverberant speech database for training speech enhancement algorithms and tts models,” 2017. [Online]. Available: http://dx.doi.org/10.7488/ds/2139

  54. [62]

    A method for the construction of minimum-redundancy codes,

    D. A. Huffman, “A method for the construction of minimum-redundancy codes,” Proc. IRE, vol. 40, no. 9, pp. 1098–1101, Sep 1952

  55. [63]

    Channel polarization: A method for constructing capacity- achieving codes,

    E. Arikan, “Channel polarization: A method for constructing capacity- achieving codes,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), July 2008, pp. 1173–1177

  56. [64]

    The jpeg still picture compression standard,

    G. Wallace, “The jpeg still picture compression standard,” IEEE Trans. Consum. Electron., vol. 38, no. 1, pp. xviii–xxxiv, Feb 1992

  57. [65]

    Low-density parity-check codes,

    R. Gallager, “Low-density parity-check codes,” IRE Trans. Inf. Theory., vol. 8, no. 1, pp. 21–28, January 1962

  58. [66]

    Bleu: a method for automatic evaluation of machine translation,

    K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, “Bleu: a method for automatic evaluation of machine translation,” in Proc. 40th Annu. Meeting Assoc. Comput. Linguistics (ACL). Philadelphia, Pennsylvania, USA: Association for Computational Linguistics, Jul. 2002, pp. 311–

  59. [67]

    Sentence-BERT: Sentence embeddings using Siamese BERT-networks,

    N. Reimers and I. Gurevych, “Sentence-BERT: Sentence embeddings using Siamese BERT-networks,” in Proc. Conf. Empir. Methods Nat. Lang. Process. 9th Int. Joint Conf. Nat. Lang. Process. (EMNLP- IJCNLP). Hong Kong, China: Association for Computational Linguistics, Nov. 2019, pp....

  60. [69]

    Alice and Bob negotiate on a large prime number P and a primitive root G of P , both of which are public global parameters

  61. [70]

    Alice generates the local private key a, computes the public key A = Ga mod P, and sends A to Bob

  62. [71]

    Bob generates the local private key b, computes the public key B = Gb mod P, and sends B to Alice

  63. [72]

    Alice receives the public key B and calculates the key I1 = Ba mod P

  64. [73]

    Bob receives the public key A and calculates the key I2 = Ab mod P

  65. [74]

    18 Alice 𝑃, 𝐺 Bob Private key 𝑎 Public key 𝐴 = 𝐺𝑎 𝑚𝑜𝑑 𝑃 𝐼1 = 𝐵𝑎 𝑚𝑜𝑑 𝑃 Private key 𝑏 Public key 𝐵 = 𝐺𝑏 𝑚𝑜𝑑 𝑃 𝐼2 = 𝐴𝑏 𝑚𝑜𝑑 𝑃 𝐴 𝐵 𝐼𝑛𝑑𝑒𝑥 = 𝐼 1 = 𝐼 2 Fig

    Because of the (Ga)b mod P= (Gb)a mod P, so Alice and Bob share the same secret number Index = I1 = I2. 18 Alice 𝑃, 𝐺 Bob Private key 𝑎 Public key 𝐴 = 𝐺𝑎 𝑚𝑜𝑑 𝑃 𝐼1 = 𝐵𝑎 𝑚𝑜𝑑 𝑃 Private key 𝑏 Public key 𝐵 = 𝐺𝑏 𝑚𝑜𝑑 𝑃 𝐼2 = 𝐴𝑏 𝑚𝑜𝑑 𝑃 𝐴 𝐵 𝐼𝑛𝑑𝑒𝑥 = 𝐼 1 = 𝐼 2 Fig. 18: Diffie-Hellman key e...

  66. [318]

    Available: https://aclanthology.org/P02-1040

    [Online]. Available: https://aclanthology.org/P02-1040

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.