Pith. sign in

REVIEW 3 major objections 4 minor 33 references

An Enhanced Privacy-preserving Federated Few-shot Learning Framework for Respiratory Disease Diagnosis

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper proposes PFFL, a federated few-shot learning framework that combines Meta-SGD with differentially private gradient noise (Meta-DPSGD) and FedAvg aggregation, claiming it can diagnose respiratory diseases from limited…

desk verdict The paper's privacy guarantee is unsupported because the DPSGD analysis does not transfer to task-level meta-gradients, and the per-disease overclaim contradicts Table 4. read the letter →

arxiv 2507.08050 v1 pith:6EKBBUHQ submitted 2025-07-10 cs.LG cs.AI

Ming Wang , Zhaoyang Duan , Dong Xue , Fangzhou Liu , Zhongheng Zhang This is my paper
classification cs.LGcs.AI
keywords federatedlearningfew-shotdifferentialprivacymeta-learningrespiratorydiseasediagnosismedicalimageclassificationmodelinversionattackMeta-SGD
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 tries to show that respiratory disease diagnosis can be done well from the small, scattered, privacy-restricted datasets that real hospitals hold. It proposes PFFL, a federated few-shot learning framework in which each client trains locally with a meta-learning algorithm, Meta-SGD, and a differentially private variant, Meta-DPSGD, adds calibrated Gaussian noise to clipped task gradients before parameters are shared. A weighted FedAvg-style aggregation combines the local models on the server. On a multi-source collection of COVID-19, SARS, MERS, and normal X-ray and CT images, the authors report accuracies above 0.8 across privacy budgets $\epsilon=1$ to $16$, with precision at $\epsilon=16$ only 1.0% below the non-private model, and large gains for data-poor clients. The paper's point is that privacy protection and few-shot federated diagnosis are compatible rather than opposing goals.

What carries the argument

The object that carries the argument is Algorithm 1, Meta-DPSGD: a meta-learning update rule that runs base learning with $\theta' = \theta - \alpha \circ \nabla_{\theta} L^{T^{tr}}(\theta)$, clips each task's meta-gradient $g^{T^{te}}$ to norm bound $C$, adds Gaussian noise $N(0,\sigma^2 C^2 I)$, and updates $(\theta,\alpha) \leftarrow (\theta,\alpha) - \beta \tilde{g}^{T^{te}}$. The privacy mechanism is the calibrated noise from the DPSGD analysis, transplanted from per-example gradients to per-task meta-gradients; the federation mechanism is the weighted averaging of uploaded parameters, giving a global model that adapts to new tasks with few samples.

What would settle it

Run a model-inversion attack on the parameters that a client uploads under Meta-DPSGD at the claimed privacy budgets: if recognizable training images are recovered, the central privacy claim is falsified. A sharper audit would recompute the sensitivity of the meta-gradient $\tilde{g}^{T^{te}}$ used in Algorithm 1 when one local example changes, and check whether the noise scale $\sigma$ satisfies the moments-accountant bound for the claimed $(\epsilon,\delta)$; the paper's Lemma 1 calibration assumes the DPSGD setting, where per-example gradients are clipped before noise, while Algorithm 1 clips per-task gradients instead.

Watch

Extended reading notes

Core claim

The central discovery is that injecting differential privacy noise into the meta-learning loop does not destroy few-shot diagnostic performance. Meta-DPSGD clips the gradient of the meta-objective for each task, averages the clipped gradients, adds Gaussian noise $N(0,\sigma^2 C^2 I)$, and then updates the meta-parameters $(\theta,\alpha)$; base-learning fine-tuning $\theta' = \theta - \alpha \circ \nabla L$ is kept noise-free. The experiments indicate that this preserves the advantages of federated learning for non-IID medical data: the framework outperforms single-modality and single-disease models on average, raises the accuracy of the smallest data-holder by as much as 59.5% compared with local-only training, and stays within about one percentage point of the unperturbed model at the largest tested privacy budget. The authors also document that federated training trails centralized training even when the total data volume matches, which they attribute to non-IID acquisition conditions across institutions.

Load-bearing premise

The load-bearing premise is that the differential privacy guarantee proven for DPSGD still holds when noise is added to clipped per-task meta-gradients rather than to per-example gradients, and that the un-noised base-learning updates do not leak information; if that transfer fails, the privacy protection claimed for Meta-DPSGD collapses.

Editorial extensions

If this is right

  • If the central claim holds, privacy-preserving few-shot federated diagnosis is practical: at $\epsilon=16$ the precision loss is only about one point, and even at $\epsilon=1$ all reported metrics remain above 0.8.
  • Federated collaboration across modalities and diseases can replace single-institution models: the CT-only model's recall on X-ray images is about 0.5, while the federated model improves cross-modality diagnosis substantially.
  • Data-poor institutions gain the most: the smallest client's accuracy rises from 0.538 to 0.858 with the framework, a 59.5% relative gain, suggesting that federated few-shot learning can mitigate resource disparity.
  • Aggregating related respiratory diseases (COVID-19, SARS, MERS) yields a more versatile diagnostic model than any single-disease model on average, with only a slight drop relative to FL without DP.

Reading between the lines

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

  • A testable extension would audit Meta-DPSGD against explicit model-inversion attacks rather than relying on the DP guarantee alone, measuring reconstruction error as a function of $\epsilon$; the paper argues protection from the theory but does not run an attack.
  • The privacy analysis transfer is the main open risk: DPSGD's moments-accountant guarantee does not automatically cover the un-noised base-learning phase, and a formal composition argument would be needed before deploying the method with a certified budget.
  • The reported non-IID gap between federated and centralized training suggests that personalized aggregation or client-specific adapters could close part of the accuracy gap; the paper stops at weighted averaging.
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 / 4 minor

Summary. The paper proposes PFFL, a federated few-shot learning framework for respiratory disease diagnosis that combines Meta-SGD with differential privacy (Meta-DPSGD) and FedAvg-based aggregation. The central claim is that Meta-DPSGD, which adds Gaussian noise to clipped task-level meta-gradients, provides a differential privacy guarantee and thereby protects against model inversion attacks. Experiments on X-ray and CT datasets evaluate the framework under centralized versus federated training, varying privacy budgets, multi-modal and multi-disease collaboration, and unbalanced data distributions, reporting accuracy, precision, recall, and F1 scores.

Significance. If the privacy guarantee were valid, the framework would be a practically relevant contribution to privacy-preserving few-shot medical image classification, where data scarcity and privacy constraints coexist. The manuscript benefits from a clearly described experimental design, the use of multiple public datasets, and the reporting of confidence intervals. However, the central scientific contribution, the Meta-DPSGD differential privacy mechanism, is not supported by a sound analysis, and this weakness is load-bearing for the paper's main claim. As a result, the empirical results, while internally mostly coherent, do not substantiate the advertised privacy-preserving property.

major comments (3)
  1. [Section 3.3, Algorithm 1; Section 2.3, Lemma 1] The differential privacy guarantee is asserted by invoking Lemma 1 from Abadi et al. [28], but that lemma applies to DP-SGD, which clips per-example gradients before averaging and noise addition. In Algorithm 1, the base-learning update theta'_i <- theta - alpha * grad_theta L_T^tr_i(theta) is performed without clipping or noise, and only the resulting task-level meta-gradient g_T^te_i is clipped and noised. Replacing one training example can change theta'_i and hence the entire meta-gradient by an amount that is not controlled by the post-hoc clipping bound C. Consequently, the global sensitivity of the released quantity is not shown to be 2C, and the condition in Lemma 1, which also relies on a sampling probability s = L/N for individual examples, has no direct analogue when batches of tasks are sampled. The reported (epsilon, delta) values in Table 2 are therefore not supported by a valid privacy analysis, and the claim in Section 3.3 that Meta-DPSGD 'can effectively safeguard against model inversion attacks' is unverified.
  2. [Section 4.4.4, Table 4] The text states that 'PFFL outperforms dedicated models in diagnosing COVID-19, SARS, and MERS', but Table 4 contradicts this for each individual disease. For COVID-19, PFFL accuracy is 0.737 versus 0.853 for the dedicated COVID-19 model (TC); for MERS, PFFL accuracy is 0.859 versus 0.906 for TM; and for SARS, PFFL accuracy is 0.878 versus 0.914 for TS. The claim appears to hold only for the average across diseases, and as written it is inaccurate. This should be corrected to avoid misrepresenting the per-disease results.
  3. [Section 4.2, Section 4.4.2] The experimental section does not report the gradient clipping bound C or the Gaussian noise scale sigma used in Meta-DPSGD, nor does it describe how these are derived from the stated privacy budgets (epsilon = 1, 2, 4, 8, 16) and delta = 10^-3. This makes the privacy-accuracy trade-off in Table 2 non-reproducible, and it is impossible to assess whether the claimed epsilon values correspond to any concrete privacy accounting. In addition, in the federated setting the noisy local parameters are aggregated and broadcast over 100 communication rounds, but the paper does not account for composition over these rounds; the epsilon reported in Table 2 cannot be interpreted as the end-to-end privacy guarantee of the federated protocol without a composition analysis.
minor comments (4)
  1. [Table 3, TX row] In Table 3, the TX row for X-ray reports precision 0.838 and recall 0.894, but an F1-score of 0.695 is inconsistent with these values; the harmonic mean is approximately 0.865. Please verify and correct this entry.
  2. [Algorithm 1] The averaging step 'g~T te <- 1/|T te_i| (sum g~T te_i)' divides by the task test-set size |T te_i|, but the summation is over tasks in the batch; the normalization should be the number of tasks. As written, the notation is ambiguous and the scaling is incorrect.
  3. [Section 2.3, Definition 3] The statement that Gaussian noise Ng = N(0, S_f^2 sigma^2) satisfies (epsilon, delta)-DP 'if and only if' delta >= (4/5) exp(-sigma^2 epsilon^2 / 2) is not a standard characterization of the Gaussian mechanism; at best it is a sufficient condition, and the constant differs from the common bound based on sigma >= sqrt(2 ln(1.25/delta)) S_f / epsilon.
  4. [Throughout] There are several typos, including 'data sacrity' in Section 1 (should be 'data scarcity') and 'Accuray' in the Table 2 header (should be 'Accuracy'). Please proofread the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the reported accuracy results are experimental outputs, Meta-DPSGD is an assembled variant of published algorithms, and the cited DP lemma is external evidence rather than a self-referential premise.

full rationale

The paper's framework is assembled from published components (Meta-SGD [18], FedAvg [25], DPSGD [28]) rather than derived from an assumption that already contains its target result. The accuracy, precision, recall, and F1 values in Tables 1-5 are empirical measurements on public datasets under fixed architectural and optimization settings; no parameter is fitted after seeing a target accuracy, and no reported diagnostic result is defined in terms of the method's own output. The central novelty, Meta-DPSGD (Algorithm 1), modifies Meta-SGD by adding clipped Gaussian noise to aggregated task-level meta-gradients; that is an algorithmic construction, not a circular definition. The privacy claim invokes Lemma 1 of Abadi et al. [28], which is external evidence rather than self-citation. The one substantive weakness is a correctness/validity gap rather than circularity: Lemma 1 in [28] is stated for per-example gradient clipping, whereas Algorithm 1 clips only after the unclipped base-learning update θ'_i = θ − α∘∇L(θ), and the paper does not show that the sensitivity and sampling-probability assumptions of Lemma 1 transfer to this setting. This is an unverified inference, not a reduction of the conclusion to its own premise. The only self-citation, [16], supports a background remark on transfer learning and is not load-bearing. Thus, under the requested standard of quoting a specific reduction of a result to its inputs, no circular step is present.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central privacy claim rests on the transfer of DPSGD's Gaussian mechanism to a task-level meta-learning setting, which is not proven. The empirical accuracy claim rests on standard supervised learning assumptions and the unstated choice of hyperparameters C and sigma.

free parameters (2)
  • Gradient clipping bound C = not reported
    Clipping threshold in Algorithm 1; controls the sensitivity of the meta-gradient and hence the DP noise scale. Set by hand, not reported.
  • Gaussian noise scale sigma = not reported
    Chosen to satisfy the privacy budget via Lemma 1 from Abadi et al.; its value is not given, so the claimed epsilon and delta cannot be verified.
assumptions (3)
  • ad hoc to paper Lemma 1 from Abadi et al. [28] applies to Meta-DPSGD's task-level sampling and per-task clipping
    The paper cites Lemma 1 to claim DP, but Algorithm 1 samples tasks and clips per-task gradients rather than per-example gradients typical of DPSGD, and the base-learning phase is unnoised. No proof is supplied for this transfer.
  • standard math The loss function is differentiable and standard backpropagation computes the required gradients.
    Implied by the use of neural networks and SGD-based optimization.
  • domain assumption Federated averaging minimizes the global loss for non-IID data (Eq. 1).
    Standard FL assumption; the paper itself notes the non-IID gap as a challenge.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Enhanced Privacy-preserving Federated Few-shot Learning Framework for Respiratory Disease Diagnosis." pith.science (2026). https://pith.science/paper/6EKBBUHQ

@misc{pith2026250708050,
  author       = {Pith},
  title        = {Pith review of: An Enhanced Privacy-preserving Federated Few-shot Learning Framework for Respiratory Disease Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6EKBBUHQ}},
  note         = {Machine review of arXiv:2507.08050}
}
read the original abstract

The labor-intensive nature of medical data annotation presents a significant challenge for respiratory disease diagnosis, resulting in a scarcity of high-quality labeled datasets in resource-constrained settings. Moreover, patient privacy concerns complicate the direct sharing of local medical data across institutions, and existing centralized data-driven approaches, which rely on amounts of available data, often compromise data privacy. This study proposes a federated few-shot learning framework with privacy-preserving mechanisms to address the issues of limited labeled data and privacy protection in diagnosing respiratory diseases. In particular, a meta-stochastic gradient descent algorithm is proposed to mitigate the overfitting problem that arises from insufficient data when employing traditional gradient descent methods for neural network training. Furthermore, to ensure data privacy against gradient leakage, differential privacy noise from a standard Gaussian distribution is integrated into the gradients during the training of private models with local data, thereby preventing the reconstruction of medical images. Given the impracticality of centralizing respiratory disease data dispersed across various medical institutions, a weighted average algorithm is employed to aggregate local diagnostic models from different clients, enhancing the adaptability of a model across diverse scenarios. Experimental results show that the proposed method yields compelling results with the implementation of differential privacy, while effectively diagnosing respiratory diseases using data from different structures, categories, and distributions.

Figures

Figures reproduced from arXiv: 2507.08050 by the authors.

Figure 1
Figure 1. The framework of PFFL. 3.3. Local model training with Meta-DPSGD Meta-DPSGD, a privacy-preserving variant of Meta-SGD, is proposed to conduct lo￾cal model training on the client to achieve privacy-preserving FSL. Analogous to Meta￾SGD, Meta-DPSGD also executes FSL through two episodes: base learning and privacy￾preserving meta-learning. The key element is that DPSGD [28] is utilized during the privacy-preserving met… view at source ↗
Figure 2
Figure 2. The results of federated training and centralized training with MAML (a) and Meta￾SGD (b) algorithms. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. The results of the proposed method with different privacy budgets. 4.4.2. Evaluating the influence of privacy budget on diagnostic effectiveness To investigate the impact of privacy budget on the proposed method, 800 COVID￾19 X-ray images and an equivalent number of normal pictures are sampled from the collected dataset, which are evenly distributed to 4 clients. The experiment is conducted with ϵ values of 1, 2, 4,… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The average performance of multi-modal data collaborative diagnosis. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: The average results of different methods diagnosing COVID-19, MERS and SARS [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: The results of the proposed method and local secure training approach using various amounts of data. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 20 canonical work pages

  1. [18]

    Meta-sgd: Learning to learn quickly for few-shot learn- ing

    Li Z, Zhou F, Chen F, et al. Meta-sgd: Learning to learn quickly for few-shot learn- ing. arXiv preprint arXiv:1707.09835 2017. https://doi.org/10.48550/arXiv. 1707.09835

    arXiv doi:10.48550/arxiv.1707.09835 2017

  2. [28]

    Deep learning with differential privacy

    Abadi M, Chu A, Goodfellow I, et al. Deep learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC conference on computer and communica- tions security. ACM, 2016 p. 308–318. https://dl.acm.org/doi/abs/10.1145/ 2976749.2978318

    arXiv 2016

  3. [1]

    Assessing the value of integrating national longitudinal shopping data into respiratory disease forecasting models

    Dolan E, Goulding J, Marshall H, et al. Assessing the value of integrating national longitudinal shopping data into respiratory disease forecasting models. Nat. Com- mun. 2023;14(1):7258. https://doi.org/10.1038/s41467-023-42776-4

    work page doi:10.1038/s41467-023-42776-4 2023

  4. [2]

    Global impact of respiratory disease: what can we do, together, to make a difference? Chest 2022;161(5):1153–1154

    Levine SM, Marciniuk DD. Global impact of respiratory disease: what can we do, together, to make a difference? Chest 2022;161(5):1153–1154. https://doi.org/ 10.1016/j.chest.2022.01.014

    work page doi:10.1016/j.chest.2022.01.014 2022

  5. [3]

    Patient exposure from radiologic and nuclear medicine procedures in the united states: procedure volume 23 and effective dose for the period 2006–2016

    Mettler Jr F A, Mahesh M, Bhargavan-Chatfield M, et al. Patient exposure from radiologic and nuclear medicine procedures in the united states: procedure volume 23 and effective dose for the period 2006–2016. Radiology 2020;295(2):418–427. https: //doi.org/10.1148/radiol.2020192256

    work page doi:10.1148/radiol.2020192256 2006

  6. [4]

    Machine-learning-based multiple ab- normality prediction with large-scale chest computed tomography volumes

    Draelos RL, Dov D, Mazurowski MA, et al. Machine-learning-based multiple ab- normality prediction with large-scale chest computed tomography volumes. Med. Image Anal. 2021;67:101857. https://doi.org/10.1016/j.media.2020.101857

    arXiv 2021

  7. [5]

    Data science and the role of artificial intelli- gence in achieving the fast diagnosis of COVID-19

    Vinod DN, Prabaharan S. Data science and the role of artificial intelli- gence in achieving the fast diagnosis of COVID-19. Chaos Solitons & Fractals 2020;140:110182. https://doi.org/10.1016/j.chaos.2020.110182

    work page arXiv 2020

  8. [6]

    Focal liver lesion diagnosis with deep learning and multistage CT imaging

    Wei Y, Yang M, Zhang M, et al. Focal liver lesion diagnosis with deep learning and multistage CT imaging. Nat. Commun. 2024;15(1):7040. https://doi.org/ 10.1038/s41467-024-51260-6

    work page doi:10.1038/s41467-024-51260-6 2024

Show all 33 references
  1. [7]

    Ultrasound-based radiomics technology in fetal lung texture analysis prediction of neonatal respiratory morbidity

    Du Y, Jiao J, Ji C, et al. Ultrasound-based radiomics technology in fetal lung texture analysis prediction of neonatal respiratory morbidity. Sci. Rep. 2022;12(1):12747. https://doi.org/10.1038/s41598-022-17129-8

    2022 doi

  2. [8]

    Evolutionary algorithms for automatic lung disease detection

    Gupta N, Gupta D, Khanna A, et al. Evolutionary algorithms for automatic lung disease detection. Measurement 2019;140:590–608. https://doi.org/10.1016/j. measurement.2019.02.042

    2019 doi

  3. [9]

    https://synapse.koreamed.org/ pdf/10.4258/hir.2016.22.4.351

    Goodfellow I Deep Learning MIT Press, 2016. https://synapse.koreamed.org/ pdf/10.4258/hir.2016.22.4.351

    2016 doi

  4. [10]

    CoroDet: a deep learning based classifica- tion for COVID-19 detection using chest X-ray images

    Hussain E, Hasan M, Rahman MA, et al. CoroDet: a deep learning based classifica- tion for COVID-19 detection using chest X-ray images. Chaos Solitons & Fractals 2021;142:110495. https://doi.org/10.1016/j.chaos.2020.110495

    2021

  5. [11]

    An efficient deep neural network framework for COVID-19 lung infection segmentation

    Jin G, Liu C, Chen X. An efficient deep neural network framework for COVID-19 lung infection segmentation. Inform. Sci. 2022;612:745–758. https://doi.org/10. 1016/j.ins.2022.08.059 . 24

    2022

  6. [12]

    Generalizing from a few examples: a survey on few- shot learning

    Wang Y, Yao Q, Kwok JT, et al. Generalizing from a few examples: a survey on few- shot learning. ACM. Comput. Surv. 2020;53(3):1–34. https://doi.org/10.1145/ 3386252

    2020

  7. [13]

    A systematic review of few-shot learning in medical imag- ing

    Pachetti E, Colantonio S. A systematic review of few-shot learning in medical imag- ing. Artif. Intell. Med. 2024;156:102949. https://doi.org/10.1016/j.artmed. 2024.102949

    2024

  8. [14]

    Few-shot learning for CT scan based COVID-19 diag- nosis

    Jiang Y, Chen H, Ko H, et al. Few-shot learning for CT scan based COVID-19 diag- nosis. In: 2021-2021 IEEE international conference on acoustics, speech and signal processing. IEEE, 2021 p. 1045–1049. https://doi.org/10.1109/ICASSP39728. 2021.9413443

    2021

  9. [15]

    Deep-COVID: Predicting COVID-19 from chest X-ray images using deep transfer learning

    Minaee S, Kafieh R, Sonka M, et al. Deep-COVID: Predicting COVID-19 from chest X-ray images using deep transfer learning. Med. Image Anal. 2020;65:101794. https://doi.org/10.1016/j.media.2020.101794

    2020

  10. [16]

    Time series modeling and forecasting of epi- demic spreading processes using deep transfer learning

    Xue D, Wang M, Liu F, et al. Time series modeling and forecasting of epi- demic spreading processes using deep transfer learning. Chaos Solitons & Fractals 2024;185:115092. https://doi.org/10.1016/j.chaos.2024.115092

    2024

  11. [17]

    MetaCOVID: a siamese neural network framework with contrastive loss for n-shot diagnosis of COVID-19 patients

    Shorfuzzaman M, Hossain MS. MetaCOVID: a siamese neural network framework with contrastive loss for n-shot diagnosis of COVID-19 patients. Pattern Recognit. 2021;113:107700. https://doi.org/10.1016/j.patcog.2020.107700

    2021

  12. [19]

    Secure, privacy-preserving and fed- erated machine learning in medical imaging

    Kaissis GA, Makowski MR, R¨ uckert D, et al. Secure, privacy-preserving and fed- erated machine learning in medical imaging. Nat. Mach. Intell. 2020;2(6):305–311. https://doi.org/10.1038/s42256-020-0186-1 . 25

    2020 doi

  13. [20]

    Federated learning for COVID-19 screening from chest X-ray images

    Feki I, Ammar S, Kessentini Y, et al. Federated learning for COVID-19 screening from chest X-ray images. Appl. Soft. Comput. 2021;106:107330. https://doi.org/ 10.1016/j.asoc.2021.107330

    2021

  14. [21]

    The secret revealer: generative model-inversion attacks against deep neural networks

    Zhang Y, Jia R, Pei H, et al. The secret revealer: generative model-inversion attacks against deep neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. IEEE, 2020 p. 253–261. https://arxiv. org/abs/1911.07135

    2020 arXiv

  15. [22]

    Model inversion attacks that exploit confidence information and basic countermeasures

    Fredrikson M, Jha S, Ristenpart T. Model inversion attacks that exploit confidence information and basic countermeasures. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security. ACM, 2015 p. 1322–1333. https://doi.org/10.1145/2810103.2813677

    2015

  16. [23]

    PPFLHE: a privacy-preserving federated learning scheme with homomorphic encryption for healthcare data

    Wang B, Li H, Guo Y, et al. PPFLHE: a privacy-preserving federated learning scheme with homomorphic encryption for healthcare data. Appl. Soft. Comput. 2023;146:110677. https://doi.org/10.1016/j.asoc.2023.110677

    2023

  17. [24]

    Federated learning-empowered disease diagnosis mech- anism in the internet of medical things: from the privacy-preservation perspec- tive

    Wang X, Hu J, Lin H, et al. Federated learning-empowered disease diagnosis mech- anism in the internet of medical things: from the privacy-preservation perspec- tive. IEEE Trans. Ind. Inf. 2022;19(7):7905–7913.https://doi.org/10.1109/TII. 2022.3210597

    2022

  18. [25]

    Communication-efficient learning of deep networks from decentralized data

    McMahan B, Moore E, Ramage D, et al. Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. PMLR, 2017 p. 1273–1282. https://proceedings.mlr.press/v54/mcmahan17a? ref=https://githubhelp.com

    2017

  19. [26]

    Our data, ourselves: privacy via dis- tributed noise generation

    Dwork C, Kenthapadi K, McSherry F, et al. Our data, ourselves: privacy via dis- tributed noise generation. In: 24th Annual international conference on the theory and applications of cryptographic techniques. Springer, 2006 p. 486–503. https: //link.springer.com/chapter/10.1007...

    2006 doi

  20. [27]

    Differential privacy for statistics: what we know and what we want to learn

    Dwork C, Smith A. Differential privacy for statistics: what we know and what we want to learn. J. Priv. Confid. 2010;1(2):135–154. https://doi.org/10.29012/ jpc.v1i2.570

    2010

  21. [29]

    Can ai help in screening viral and COVID-19 pneumonia? IEEE Access 2020;8:132665–132676.https://doi.org/10

    Chowdhury ME, Rahman T, Khandakar A, et al. Can ai help in screening viral and COVID-19 pneumonia? IEEE Access 2020;8:132665–132676.https://doi.org/10. 1109/ACCESS.2020.3010287

    2020

  22. [30]

    Exploring the effect of image enhance- ment techniques on COVID-19 detection using chest X-ray images

    Rahman T, Khandakar A, Qiblawey Y, et al. Exploring the effect of image enhance- ment techniques on COVID-19 detection using chest X-ray images. Comput. Biol. Med. 2021;132:104319. https://doi.org/10.1016/j.compbiomed.2021.104319

    2021

  23. [31]

    Deep learning for reliable classifica- tion of COVID-19, MERS, and SARS from chest X-ray images

    Tahir AM, Qiblawey Y, Khandakar A, et al. Deep learning for reliable classifica- tion of COVID-19, MERS, and SARS from chest X-ray images. Cogn. Comput. 2022;14:1–21. https://doi.org/10.1007/s12559-021-09955-1

    2022 doi

  24. [32]

    COVID19-CT-dataset: a CT scan dataset about COVID-19

    Yang X, He X, Zhao J, et al. COVID19-CT-dataset: a CT scan dataset about COVID-19. arXiv preprint arXiv:2003.13865 2020. https://doi.org/10.48550/ arXiv.2003.13865

    2003 doi arXiv

  25. [33]

    Model-agnostic meta-learning for fast adaptation of deep networks

    Finn C, Abbeel P, Levine S. Model-agnostic meta-learning for fast adaptation of deep networks. In: International conference on machine learning. PMLR, 2017 p. 1126–1135. https://proceedings.mlr.press/v70/finn17a.html . 27

    2017

Pith tools

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