pith. sign in

arxiv: 2606.07640 · v1 · pith:BB3PR6Q2new · submitted 2026-06-01 · 💻 cs.CV · cs.AI· cs.LG

No Free Lunch for Synthetic Images under Data Scarcity Conditions

Borja Arroyo Galende , Alejandro Almod\'ovar , Patricia A. Apell\'aniz , Juan Parras , Silvia Uribe , Santiago Zazo This is my paper

Pith reviewed 2026-06-28 14:50 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LG
keywords synthetic datadifferential privacygenerative modelsGANDDPMVAEimage fidelitydata utility
0
0 comments X

The pith

GANs and DDPMs maintain higher fidelity and utility than VAEs when differential privacy noise is added during training on scarce image data.

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

The paper studies trade-offs among image fidelity, privacy protection, and downstream task utility when generating synthetic images from limited datasets under privacy constraints. It applies a single framework that measures all three aspects together to three standard generative models across general and medical image collections. Results show that GANs and diffusion models degrade less than variational autoencoders as privacy noise increases. This matters for applications that must produce usable synthetic data without leaking private information. The work concludes that model choice matters once privacy mechanisms enter the training process.

Core claim

When differential privacy mechanisms are introduced during training, GAN and DDPM maintain higher fidelity and downstream utility across a range of noise levels, while VAE degrades more rapidly as privacy constraints increase.

What carries the argument

Joint evaluation framework that simultaneously measures fidelity, privacy, and utility of synthetic images produced by generative models.

If this is right

  • GANs and DDPMs retain usable downstream performance longer than VAEs once privacy noise is introduced.
  • No generative model achieves high scores on all three dimensions at once under strong privacy constraints.
  • Medical imaging datasets show the same model ordering as general-purpose ones.
  • Evaluation must track fidelity, privacy, and utility jointly rather than separately.

Where Pith is reading between the lines

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

  • The observed robustness ordering could guide selection of generators for other privacy-sensitive domains such as tabular or time-series data.
  • If the framework were applied to larger-scale models, the relative degradation rates might change.
  • Alternative privacy mechanisms beyond the tested noise levels might alter which model appears most stable.

Load-bearing premise

The chosen metrics and the three image datasets together capture the fidelity-privacy-utility trade-offs without systematic bias.

What would settle it

Re-running the same models and privacy levels on a new dataset where VAE fidelity and utility remain equal to or above those of GAN and DDPM at high noise would contradict the robustness ordering.

Figures

Figures reproduced from arXiv: 2606.07640 by Alejandro Almod\'ovar, Borja Arroyo Galende, Juan Parras, Patricia A. Apell\'aniz, Santiago Zazo, Silvia Uribe.

Figure 1
Figure 1. Figure 1: Privacy estimation framework. X is the original dataset; p(X) its distribution; xq is the query sample excluded in the ablated model; Xs and X′ s are synthetic datasets generated by the victim (θv) and ablated (θv) models, respectively. A large sample size is used to estimate p(x ∗ q ) under both models, enabling calculation of ε by using Equation (2). 3.3 Fidelity Evaluation Fidelity evaluates how closely… view at source ↗
Figure 2
Figure 2. Figure 2: Design for the fidelity analysis. X is the data; p(X) its distribution; Xt and Xh are samples from p(X); Xs is the synthetic data set. The sizes of Xh and Xs are matched for fair metric estimation. rics, such as the Jensen-Shannon distance, may also be employed [3]. These metrics as￾sess statistical similarity between distributions and are particularly suitable when working with latent representations or w… view at source ↗
Figure 3
Figure 3. Figure 3: Design for the utility analysis. X is the data; p(X) its distribution; (Xt , yt) and (Xh, yh) are real labeled data; Xs is synthetic. The synthetic and holdout sets have the same size to fairly assess task performance. Because data scarcity significantly affects task performance, we limit our analysis to settings that enforce the same sample budget between experiments. Consequently, com￾parisons are made o… view at source ↗
Figure 4
Figure 4. Figure 4: Spearman rank correlation coefficients across dimensionality reduction methods [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: MNIST projected into a 2D space through the Isomap model and represented [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: VAE image reconstruction at different DPSGD noise levels. The upper row [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Generated samples from all models under different DPSGD noise levels. The [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Projections and 2D histograms for three models at [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Isomap projection for the GAN model at σ = 0.03 covering all modes from the original data. In contrast to noiseless training, DPSGD shows a regularising effect which alleviates mode collapse. 4.2.3 Quantitative Analysis [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparative results in MNIST measured through FID, IS, empirical [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: KS distance and normalised area between ECDF curves for several [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: GradCAM visualisations over real data from a CNN classifier trained on [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Isomap projections for the OCTMNIST data set with the corresponding labels. [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Isomap projections for the OrganAMNIST data set with the corresponding [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: KS distance and normalised area between ECDF curves for several [PITH_FULL_IMAGE:figures/full_fig_p020_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Comparative results in OCTMNIST across FID, IS, empirical [PITH_FULL_IMAGE:figures/full_fig_p021_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: KS distance and normalised area between ECDF curves for several [PITH_FULL_IMAGE:figures/full_fig_p022_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Comparative results in OrganAMNIST across FID, IS, empirical [PITH_FULL_IMAGE:figures/full_fig_p023_18.png] view at source ↗
read the original abstract

This study investigates the trade-offs between fidelity, privacy, and utility in synthetic data generation under conditions of data scarcity and privacy sensitivity. We propose an evaluation framework that jointly assesses these three dimensions and apply it to three widely used generative models, VAE, GAN, and DDPM. The evaluation spans three image datasets, MNIST, OCTMNIST, and OrganAMNIST, encompassing both general-purpose and medical imaging domains. Notable differences arise between the three models in their behaviour when differential privacy mechanisms are introduced during training. GAN and DDPM demonstrate greater robustness, maintaining higher fidelity and downstream utility across a range of noise levels, while VAE degrades more rapidly as privacy constraints increase. This study highlights the importance of a multidimensional evaluation of deep generative models, also noting that their behaviour significantly differs when privacy techniques are applied.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper investigates trade-offs between fidelity, privacy, and utility in synthetic image generation under data scarcity and privacy sensitivity. It proposes a joint evaluation framework applied to VAE, GAN, and DDPM across MNIST, OCTMNIST, and OrganAMNIST datasets. The central empirical claim is that GAN and DDPM exhibit greater robustness to differential privacy mechanisms during training, preserving higher fidelity and downstream task utility across noise levels, whereas VAE degrades more rapidly as privacy constraints tighten.

Significance. If the empirical results hold under rigorous controls, the work provides a useful multidimensional lens for selecting generative models in privacy-sensitive, data-scarce regimes such as medical imaging. The comparative robustness finding across model families is potentially actionable for practitioners, though its impact hinges on the transparency and reproducibility of the evaluation protocol, metric definitions, and statistical validation.

major comments (2)
  1. [Abstract / Methods (implied)] The abstract and provided description supply no concrete details on the differential privacy implementation (e.g., noise scale, clipping norms, or accountant method), the exact fidelity/utility metrics, statistical tests, or train/test splits. Without these, the load-bearing claim that GAN/DDPM are “more robust” cannot be verified or reproduced; this directly affects the central comparative result.
  2. [Evaluation framework (implied)] The joint evaluation framework is presented as comprehensive, yet no evidence is given that the chosen metrics are free of hidden biases or that they jointly capture the fidelity-privacy-utility space without trade-off artifacts. This assumption underpins the entire comparative analysis.
minor comments (2)
  1. [Abstract] Clarify whether the reported “noise levels” correspond to a single privacy budget ε or to multiple mechanisms; inconsistent terminology risks reader confusion.
  2. [Datasets (implied)] The datasets span general and medical domains; a brief justification for their selection and any domain-specific preprocessing would strengthen the generalizability claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on reproducibility and evaluation rigor. We address each major comment below and will revise the manuscript to strengthen transparency without altering the core empirical findings.

read point-by-point responses

  1. Referee: [Abstract / Methods (implied)] The abstract and provided description supply no concrete details on the differential privacy implementation (e.g., noise scale, clipping norms, or accountant method), the exact fidelity/utility metrics, statistical tests, or train/test splits. Without these, the load-bearing claim that GAN/DDPM are “more robust” cannot be verified or reproduced; this directly affects the central comparative result.

    Authors: We agree that the abstract is intentionally concise and that the provided high-level description omits implementation specifics. The full manuscript contains these details in the Methods and Experimental Setup sections (including DP-SGD parameters via Opacus, accountant method, clipping norms, noise scales, FID/SSIM for fidelity, downstream accuracy for utility, and explicit train/test splits on MNIST/OCTMNIST/OrganAMNIST). To directly address verifiability, we will add a consolidated 'Reproducibility' subsection (or appendix table) listing all hyperparameters, metric formulas, and statistical procedures (e.g., 5-run averages with standard deviations). This revision will make the robustness claims fully reproducible. revision: yes

  2. Referee: [Evaluation framework (implied)] The joint evaluation framework is presented as comprehensive, yet no evidence is given that the chosen metrics are free of hidden biases or that they jointly capture the fidelity-privacy-utility space without trade-off artifacts. This assumption underpins the entire comparative analysis.

    Authors: The framework combines established metrics (FID and perceptual similarity for fidelity, epsilon-bounded DP for privacy, and task-specific accuracy for utility) drawn from prior literature on generative model evaluation under privacy constraints. While we do not claim the metrics are entirely bias-free (no finite set of metrics can be), we will expand the Evaluation Framework section with a short justification paragraph citing supporting references and include a brief sensitivity discussion on potential artifacts. If the referee has specific alternative metrics in mind, we are open to incorporating them as additional experiments. revision: partial

Circularity Check

0 steps flagged

Empirical study with no derivations or self-referential predictions

full rationale

The manuscript is a comparative empirical study applying an evaluation framework to VAE, GAN, and DDPM on three image datasets under differential privacy. No equations, derivations, fitted parameters relabeled as predictions, or load-bearing self-citations appear in the provided text or abstract. Claims rest on direct experimental outcomes rather than any chain that reduces to its own inputs by construction. The evaluation framework is introduced and applied without internal circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities. The evaluation implicitly assumes standard ML dataset representativeness and metric validity for fidelity, privacy, and utility.

pith-pipeline@v0.9.1-grok · 5689 in / 1035 out tokens · 19543 ms · 2026-06-28T14:50:03.974063+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

55 extracted references · 1 linked inside Pith

  1. [1]

    Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang

    Martín Abadi, Andy Chu, Ian Goodfellow, H. Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep Learning with Differential Privacy. https://arxiv.org/abs/1607.00133v2, July 2016

    Pith/arXiv arXiv 2016

  2. [2]

    On the fidelity versus privacy and utility trade-off of synthetic patient data.iScience, 28(5):112382, May 2025

    Tim Adams, Colin Birkenbihl, Karen Otte, Hwei Geok Ng, Jonas Adrian Rieling, Anatol-Fiete Näher, Ulrich Sax, Fabian Prasser, and Holger Fröhlich. On the fidelity versus privacy and utility trade-off of synthetic patient data.iScience, 28(5):112382, May 2025

    2025

  3. [3]

    Apellániz, Ana Jiménez, Borja Arroyo Galende, Juan Parras, and Santi- ago Zazo

    Patricia A. Apellániz, Ana Jiménez, Borja Arroyo Galende, Juan Parras, and Santi- ago Zazo. Synthetic Tabular Data Validation: A Divergence-Based Approach.IEEE Access, 12:103895–103907, 2024

    2024

  4. [4]

    Schwartz

    Mukund Balasubramanian and Eric L. Schwartz. The Isomap Algorithm and Topo- logical Stability.Science, 295(5552):7–7, January 2002

    2002

  5. [5]

    A systematic review on overfitting control in shallow and deep neural networks.Artif

    Mohammad Mahdi Bejani and Mehdi Ghatee. A systematic review on overfitting control in shallow and deep neural networks.Artif. Intell. Rev., 54(8):6391–6438, December 2021

    2021

  6. [6]

    A mathematical theory of adaptive control processes.Proceedings of the National Academy of Sciences, 45(8):1288–1290, August 1959

    Richard Bellman and Robert Kalaba. A mathematical theory of adaptive control processes.Proceedings of the National Academy of Sciences, 45(8):1288–1290, August 1959. 26

    1959

  7. [7]

    Representation Learning: A Review and New Perspectives.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798–1828, August 2013

    Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation Learning: A Review and New Perspectives.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798–1828, August 2013

    2013

  8. [8]

    Willcocks

    Sam Bond-Taylor, Adam Leach, Yang Long, and Chris G. Willcocks. Deep Gen- erative Modelling: A Comparative Review of VAEs, GANs, Normalizing Flows, Energy-Based and Autoregressive Models.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7327–7347, November 2022

    2022

  9. [9]

    Extracting Training Data from Large Language Models

    Nicholas Carlini, Florian Tramèr, Eric Wallace, Matthew Jagielski, Ariel Herbert- Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Úlfar Erlingsson, Alina Oprea, and Colin Raffel. Extracting Training Data from Large Language Models. In30th USENIX Security Symposium (USENIX Security 21), pages 2633–

  10. [10]

    USENIX Association, August 2021

    2021

  11. [11]

    Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sen- gupta, and Anil A. Bharath. Generative Adversarial Networks: An Overview.IEEE Signal Processing Magazine, 35(1):53–65, January 2018

    2018

  12. [12]

    Saverio D’Amico, Daniele Dall’Olio, Claudia Sala, Lorenzo Dall’Olio, Elisabetta Sauta, Matteo Zampini, Gianluca Asti, Luca Lanino, Giulia Maggioni, Alessia Campagna, Marta Ubezio, Antonio Russo, Maria Elena Bicchieri, Elena Riva, Cristina A. Tentori, Erica Travaglino, Pierandrea Morandini, Victor Savevski, Ar- mando Santoro, Iñigo Prada-Luengo, Anders Kro...

    2023

  13. [13]

    The mnist database of handwritten digit images for machine learning research.IEEE Signal Processing Magazine, 29(6):141–142, 2012

    Li Deng. The mnist database of handwritten digit images for machine learning research.IEEE Signal Processing Magazine, 29(6):141–142, 2012

    2012

  14. [14]

    Regulation (EU) 2025/327 of the European Parliament and of the Council on the European Health Data Space

    Directorate-General for Health and Food Safety (DG SANTE), European Com- mission. Regulation (EU) 2025/327 of the European Parliament and of the Council on the European Health Data Space. https://eur-lex.europa.eu/legal- content/EN/TXT/?uri=CELEX

    2025

  15. [15]

    DifferentiallyPrivate Diffusion Models.Transactions on Machine Learning Research, 2:1–25, May 2023

    TimDockhorn, TianshiCao, ArashVahdat, andKarstenKreis. DifferentiallyPrivate Diffusion Models.Transactions on Machine Learning Research, 2:1–25, May 2023

    2023

  16. [16]

    Gökcen Eraslan, Žiga Avsec, Julien Gagneur, and Fabian J. Theis. Deep learning: New computational modelling techniques for genomics.Nature Reviews Genetics, 20(7):389–403, July 2019

    2019

  17. [17]

    prentice hall professional technical reference, 2002

    David A Forsyth and Jean Ponce.Computer Vision: A Modern Approach. prentice hall professional technical reference, 2002

    2002

  18. [18]

    Apellániz, Juan Parras, Santiago Zazo, and Silvia Uribe

    Borja Arroyo Galende, Patricia A. Apellániz, Juan Parras, Santiago Zazo, and Silvia Uribe. Membership Inference Attacks and Differential Privacy: A study within the context of Generative Models.IEEE Open Journal of the Computer Society, pages 1–10, 2025. 27

    2025

  19. [19]

    Mauro Giuffrè and Dennis L. Shung. Harnessing the power of synthetic data in healthcare: Innovation, application, and privacy.npj Digital Medicine, 6(1):1–8, October 2023

    2023

  20. [20]

    Aldren Gonzales, Guruprabha Guruswamy, and Scott R. Smith. Synthetic data in health care: A narrative review.PLOS Digital Health, 2(1):e0000082, January 2023

    2023

  21. [21]

    The MIT Press, October 2016

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep Learning. The MIT Press, October 2016

    2016

  22. [22]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139–144, October 2020

    2020

  23. [23]

    A Grey-box Attack against Latent Diffusion Model-based Image Editing by Posterior Collapse, February 2025

    Zhongliang Guo, Chun Tong Lei, Lei Fang, Shuai Zhao, Yifei Qian, Jingyu Lin, Zeyu Wang, Cunjian Chen, Ognjen Arandjelović, and Chun Pong Lau. A Grey-box Attack against Latent Diffusion Model-based Image Editing by Posterior Collapse, February 2025

    2025

  24. [24]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InProceedings of the 34th International Conference on Neural Information Process- ing Systems, NIPS ’20, pages 6840–6851, Red Hook, NY, USA, December 2020. Curran Associates Inc

    2020

  25. [25]

    Cohen, and Adrian Weller

    James Jordon, Lukasz Szpruch, Florimond Houssiau, Mirko Bottarelli, Giovanni Cherubin, Carsten Maple, Samuel N. Cohen, and Adrian Weller. Synthetic Data – what, why and how?, 2024

    2024

  26. [26]

    A scoping re- view of privacy and utility metrics in medical synthetic data.NPJ Digital Medicine, 8:60, January 2025

    Bayrem Kaabachi, Jérémie Despraz, Thierry Meurers, Karen Otte, Mehmed Halilovic, Bogdan Kulynych, Fabian Prasser, and Jean Louis Raisaro. A scoping re- view of privacy and utility metrics in medical synthetic data.NPJ Digital Medicine, 8:60, January 2025

    2025

  27. [27]

    Theoretical Analysis of Density Ratio Estimation.IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, E93-A(4):787–798, 2010

    Takafumi Kanamori, Taiji Suzuki, and Masashi Sugiyama. Theoretical Analysis of Density Ratio Estimation.IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, E93-A(4):787–798, 2010

    2010

  28. [28]

    No free lunch in data privacy

    Daniel Kifer and Ashwin Machanavajjhala. No free lunch in data privacy. InPro- ceedings of the 2011 ACM SIGMOD International Conference on Management of Data, SIGMOD ’11, pages 193–204, New York, NY, USA, June 2011. Association for Computing Machinery

    2011

  29. [29]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes. InPro- ceedings of the 2nd International Conference on Learning Representations (ICLR), 2014

    2014

  30. [30]

    Lautrup, Tobias Hyrup, Arthur Zimek, and Peter Schneider-Kamp

    Anton D. Lautrup, Tobias Hyrup, Arthur Zimek, and Peter Schneider-Kamp. Syn- theval: A framework for detailed utility and privacy evaluation of tabular synthetic data.Data Mining and Knowledge Discovery, 39(1):1–6, December 2024

    2024

  31. [31]

    Convolutional networks for images, speech, and time series

    Yann LeCun and Yoshua Bengio. Convolutional networks for images, speech, and time series. InThe Handbook of Brain Theory and Neural Networks, pages 255–258. MIT Press, Cambridge, MA, USA, October 1998. 28

    1998

  32. [32]

    Cresswell, and Anthony L

    Gabriel Loaiza-Ganem, Brendan Leigh Ross, Jesse C. Cresswell, and Anthony L. Caterini. Diagnosing and Fixing Manifold Overfitting in Deep Generative Models. Transactions on Machine Learning Research, 1:1–34, January 2022

    2022

  33. [33]

    Machine Learning for Synthetic Data Generation: A Review, June 2024

    Yingzhou Lu, Minjie Shen, Huazheng Wang, Xiao Wang, Capucine van Rechem, Tianfan Fu, and Wenqi Wei. Machine Learning for Synthetic Data Generation: A Review, June 2024

    2024

  34. [34]

    Jun Lv, Jin Zhu, and Guang Yang. Which GAN? A comparative study of gen- erative adversarial network-based fast MRI reconstruction.Philosophical Trans- actions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 379(2200):20200203, May 2021

    2021

  35. [35]

    P Mack and M Rosenblatt

    Y. P Mack and M Rosenblatt. Multivariatek-nearest neighbor density estimates. Journal of Multivariate Analysis, 9(1):1–15, March 1979

    1979

  36. [36]

    UMAP: Uni- form Manifold Approximation and Projection.Journal of Open Source Software, 3(29):861, 2018

    Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. UMAP: Uni- form Manifold Approximation and Projection.Journal of Open Source Software, 3(29):861, 2018

    2018

  37. [37]

    Mendes, Aziz Barbar, and Marwa Refaie

    Jorge M. Mendes, Aziz Barbar, and Marwa Refaie. Synthetic data generation: A privacy-preserving approach to accelerate rare disease research.Frontiers in Digital Health, 7, March 2025

    2025

  38. [38]

    Rashidi, Samer Albahra, Brian P

    Hooman H. Rashidi, Samer Albahra, Brian P. Rubin, and Bo Hu. A novel and fully automated platform for synthetic tabular data generation and validation.Scientific Reports, 14(1):23312, October 2024

    2024

  39. [39]

    VariationalInferencewithNormalizingFlows

    DaniloRezendeandShakirMohamed. VariationalInferencewithNormalizingFlows. InProceedings of the 32nd International Conference on Machine Learning, pages 1530–1538. PMLR, June 2015

    2015

  40. [40]

    A Meta-Analysis of Overfitting in Machine Learning

    Rebecca Roelofs, Vaishaal Shankar, Benjamin Recht, Sara Fridovich-Keil, Moritz Hardt, John Miller, and Ludwig Schmidt. A Meta-Analysis of Overfitting in Machine Learning. InAdvances in Neural Information Processing Systems, volume 32, pages 14854–14864. Curran Associates, Inc., December 2019

    2019

  41. [41]

    U-Net: Convolutional Net- works for Biomedical Image Segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Net- works for Biomedical Image Segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells, and Alejandro F. Frangi, editors,Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, pages 234–241, Cham, 2015. Springer International Publishing

    2015

  42. [42]

    An introduction to deep generative modeling

    Lars Ruthotto and Eldad Haber. An introduction to deep generative modeling. GAMM-Mitteilungen, 44(2):e202100008, 2021

    2021

  43. [43]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedan- tam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedan- tam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. In2017 IEEE International Conference on Computer Vision (ICCV), pages 618–626, October 2017. 29

    2017

  44. [44]

    Benchmark- ing Synthetic Tabular Data: A Multi-Dimensional Evaluation Framework, April 2025

    Andrey Sidorenko, Michael Platzer, Mario Scriminaci, and Paul Tiwald. Benchmark- ing Synthetic Tabular Data: A Multi-Dimensional Evaluation Framework, April 2025

    2025

  45. [45]

    Shuang Song, Kamalika Chaudhuri, and Anand D. Sarwate. Stochastic gradient descent with differentially private updates. In2013 IEEE Global Conference on Signal and Information Processing, pages 245–248, December 2013

    2013

  46. [46]

    Differ- entially Private Variational Autoencoders with Term-wise Gradient Aggregation

    Tsubasa Takahashi, Shun Takagi, Hajime Ono, and Tatsuya Komatsu. Differ- entially Private Variational Autoencoders with Term-wise Gradient Aggregation. https://arxiv.org/abs/2006.11204v1, June 2020

    arXiv 2006

  47. [47]

    PreventingoversmoothinginVAEviageneralizedvariance parameterization.Neurocomputing, 509:137–156, October 2022

    Yuhta Takida, Wei-Hsiang Liao, Chieh-Hsin Lai, Toshimitsu Uesaka, Shusuke Taka- hashi, andYukiMitsufuji. PreventingoversmoothinginVAEviageneralizedvariance parameterization.Neurocomputing, 509:137–156, October 2022

    2022

  48. [48]

    Terrell and David W

    George R. Terrell and David W. Scott. Variable Kernel Density Estimation.The Annals of Statistics, 20(3):1236–1265, 1992

    1992

  49. [49]

    Tsirikoglou, G

    A. Tsirikoglou, G. Eilertsen, and J. Unger. A Survey of Image Synthesis Methods for Visual Machine Learning.Computer Graphics Forum, 39(6):426–451, 2020

    2020

  50. [50]

    Department of Health and Human Services, Office for Civil Rights

    U.S. Department of Health and Human Services, Office for Civil Rights. Notice of Proposed Rulemaking to Update the HIPAA Security Rule. https://www.federalregister.gov/documents/2025/01/01/2025-00001/notice-of- proposed-rulemaking-to-update-the-hipaa-security-rule, January 2025

    2025

  51. [51]

    VisualizingDatausingt-SNE.Journal of Machine Learning Research, 9(86):2579–2605, 2008

    LaurensvanderMaatenandGeoffreyHinton. VisualizingDatausingt-SNE.Journal of Machine Learning Research, 9(86):2579–2605, 2008

    2008

  52. [52]

    VariableKernelDensityEstimationin High-Dimensional Feature Spaces.Proceedings of the AAAI Conference on Artificial Intelligence, 31(1), February 2017

    ChristiaanvanderWaltandEtienneBarnard. VariableKernelDensityEstimationin High-Dimensional Feature Spaces.Proceedings of the AAAI Conference on Artificial Intelligence, 31(1), February 2017

    2017

  53. [53]

    MedMNIST v2 - A large-scale lightweight benchmark for 2D and 3D biomedical image classification.Scientific Data, 10(1):41, January 2023

    Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. MedMNIST v2 - A large-scale lightweight benchmark for 2D and 3D biomedical image classification.Scientific Data, 10(1):41, January 2023

    2023

  54. [54]

    No Free Lunch Theorem for Security and Utility in Federated Learning.ACM Trans

    Xiaojin Zhang, Hanlin Gu, Lixin Fan, Kai Chen, and Qiang Yang. No Free Lunch Theorem for Security and Utility in Federated Learning.ACM Trans. Intell. Syst. Technol., 14(1):1:1–1:35, November 2022

    2022

  55. [55]

    GenerativeModeling.Open Encyclopedia of Cognitive Science, July 2024

    Jun-YanZhuandPhillipIsola. GenerativeModeling.Open Encyclopedia of Cognitive Science, July 2024. 30

    2024