acceptodds
Under review as a conference paper at ICLR 2027

DPLongSyn: Differentially Private Longitudinal Image Synthesis

Abstract

Longitudinal image data capture changes in the same subject over time, such as anatomical changes across medical scans of the same patient. Such data often contain sensitive individual information, motivating synthetic data release with formal differential privacy (DP) guarantees. However, existing DP image synthesis methods mainly focus on individual images, while longitudinal synthesis also requires coherent changes between observations that agree with the associated records. To this end, we introduce DPLongSyn, a DP longitudinal image synthesis framework that learns record-conditioned longitudinal transitions and uses them to synthesize follow-up images from public source images. DPLongSyn first learns a transition prior from public longitudinal pairs. It then uses private data to generate DP synthetic records describing subject attributes and longitudinal changes and to adapt the transition model using DP-SGD. At synthesis time, the adapted model generates a deformation conditioned on a public source image and a DP synthetic record, then applies it to produce the follow-up image. Experiments on three benchmarks show that DPLongSyn achieves better synthetic data quality, longitudinal consistency, and downstream utility than DP synthesis baselines. To the best of our knowledge, this is the first work to study DP longitudinal image synthesis.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential Privacy, Longitudinal Image Synthesis, Synthetic Data

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ctS7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies differentially private synthesis of longitudinal image pairs and associated structured records. DPLongSyn builds on a publicly trained deformation-based generative framework, combines DP record synthesis with parameter-efficient private adaptation of the transition model, and produces follow-up images by applying record-conditioned deformations to public source images. Experiments cover two brain MRI benchmarks and one facial-expression benchmark at multiple privacy budgets, evaluating image and record fidelity, longitudinal consistency, downstream classification, and membership inference.

Critical strengths
  1. The paper addresses a meaningful problem with a well-motivated design. Longitudinal synthesis requires generated observations to exhibit changes that agree with their associated records, beyond the realism of individual images. Using public data to supply source appearance and a transition prior, while concentrating private learning on records and conditional transitions, is a sensible approach to this requirement. The separation of the two private mechanisms from subsequent public-data post-processing also makes the end-to-end privacy reasoning clear.

  2. The reported experiments provide encouraging comparative evidence across two domains. DPLongSyn achieves higher record–transition correlation than the evaluated baselines across the reported dataset–budget settings, as well as the highest mean follow-up AUC in Table 2 and progression AUC in Table 3. At ε=8\varepsilon=8, for example, the follow-up AUC on CelebV-HQ improves from 0.503 for the strongest evaluated baseline to 0.608, and the ADNI progression AUC improves from 0.522 to 0.559. Training classifiers on synthetic releases and testing them on held-out real data provides evidence of utility beyond distributional image metrics. The evaluation across multiple classifier families is also useful.

  3. The controls and ablations support specific aspects of the method and expose relevant tradeoffs. Figure 3 shows improved record–transition consistency after private adaptation, an advantage for updating normalization parameters together with the conditioning MLP over the tested alternatives, and a tradeoff between record fidelity and transition consistency under a fixed composed privacy budget. Table 25 holds the DP records and selected public sources fixed when comparing transition models, and appropriately reports that private adaptation does not consistently improve FID or follow-up classification. Table 7 additionally provides explicit numerical privacy accounting for DPLongSyn.

Critical weaknesses
  1. The adapted baselines’ private-training configurations are not specified in enough detail to fully reproduce and assess the comparison. Appendix B.5 identifies the public pretraining cohorts, target privacy levels, sampling rate, training steps, and several adaptations. However, the submission does not report the adapted neural models’ total and privately trainable parameter counts or clearly list their private update scopes. It also does not provide baseline numerical accounting tables comparable to Table 7, including the applicable noise multipliers, clipping norms, and mechanism-level budget allocations. These omissions limit independent verification of the reported common privacy setting and interpretation of how the private optimization strategies differ. This is particularly relevant because Figure 3(b) demonstrates sensitivity to the choice of trainable parameters. Different parameter counts do not by themselves invalidate comparisons between complete methods, but the actual configurations should be available.

  2. The basis for selecting the baseline training schedules and checkpoints is insufficiently documented. Appendix B.5 specifies a DP-SGD sampling rate of 0.5 and 300 optimization steps for the smaller datasets, whereas DPLongSyn uses a different schedule in Table 7. Different schedules can be appropriate and need not be matched mechanically. However, the paper does not provide the validation or convergence evidence motivating the baseline choices, nor does it clearly specify the corresponding public-pretraining and private-training checkpoint-selection procedures. Appendix B.4 gives a public-validation criterion for DPLongSyn, but the baseline description does not provide equivalent detail. This leaves uncertainty about whether the adapted methods received training configurations appropriate to the changed resolution and dataset sizes.

Questions for authors
  1. Model capacity and private update scope (Weakness 1): Please report the total and privately trainable parameter counts for DPLongSyn and each adapted neural baseline component, and specify which modules are updated or frozen during private training after public pretraining.

  2. Baseline privacy calibration (Weakness 1): Could you provide numerical accounting settings for the adapted baselines, analogous to Table 7? Please include clipping norms and noise multipliers for DP-SGD components, mechanism-level budget allocations, and the composed privacy budgets at the stated δ\delta , and clarify whether the implemented sampling procedures match the respective accountants’ assumptions.

  3. Training schedules and checkpoint selection (Weakness 2): Please explain how the baseline public-pretraining and private-training horizons and checkpoints were selected, and provide any existing validation or convergence evidence supporting the sampling rate and 300-step private-training schedule in Appendix B.5.

  4. Explanation of high FID values: The absolute FID values in Table 1 are relatively high. Could you discuss the likely reasons for these values in your experimental setting and what they imply about the fidelity of the generated follow-up images? Please also explain how the high FID of the zero-deformation public-image control in Table 25 informs this interpretation.

4 comments
AuthorsOpenReview ↗

Responses to Reviewer ctS7

We thank the reviewer for the careful reading and thoughtful comments. We address the reviewer's questions in the following and have revised the paper accordingly. The changes are marked in blue in our revision. We hope the responses below address the reviewer's concerns.

W1, Q1, and Q2: The baselines’ private training configurations and numerical privacy accounting.

We have expanded Appendix B.5 to report the private update scopes and numerical privacy accounting for all baselines. Below, we provide these configurations for ADNI at ε=8\varepsilon=8 and δ=10−5\delta=10^{-5}.

MethodNeural componentTotal parametersUpdated by DP-SGDPrivate update scope
DPLongSynDiffusion U-Net3,547,16829,408 (0.83%)Normalization layer and conditioning MLP. All other U-Net parameters and the public registration decoder are frozen.
DP-IndepImage network2,474,7872,474,787 (100%)Full image network. The tabular stage uses AIM.
DP-TabImageConditional image network2,592,9312,592,931 (100%)Full conditional image network. The tabular stage uses AIM.
DP-ImageTabImage network2,474,7872,474,787 (100%)Full image network.
DP-ImageTabTabular network532,333532,333 (100%)Full tabular network.
DP-JointDMJoint image–record network2,229,6802,229,680 (100%)Full joint network.
DP-JointMERFGenerator4,328,012N/A (no DP-SGD)The generator is trained on privatized label statistics and random-feature means.
MethodPrivacy budget allocationClip CCNoise parameter σ\sigmaqq / updatesComposed ε\varepsilon
DPLongSynRecord ε=4\varepsilon=4; remaining budget composed with image stage0.0522.8307 (image)0.05 / 5,5408.000
DP-Indep20% zCDP to AIM; remaining to image1.01.4258 (AIM), 6.25 (image)0.5 / 3007.990
DP-TabImage20% zCDP to AIM; remaining to image1.01.4258 (AIM), 6.25 (image)0.5 / 3007.990
DP-ImageTabImage and tabular stages composed1.011.637 (image), 6.348 (tabular)0.5 / 300 per stage7.999
DP-JointDMFull budget to joint DP-SGD1.05.6010.5 / 3007.995
DP-JointMERF5% zCDP to labels; 95% to embeddingsN/AN/A (no DP-SGD)N/A8.000

All DP-SGD stages use Poisson sampling, consistent with the assumptions of the corresponding privacy accountants. We also provide the corresponding accounting details for all three datasets at ε∈2,8,32\varepsilon\in\\{2,8,32\\} in Appendix B.5.

W2 and Q3: Training schedules and checkpoint selection.

Training schedule. For the four baselines, we follow the original papers' 150 training epochs. However, using a sampling rate of q=0.05q=0.05 on our smaller private datasets results in small batches and noisy updates. We therefore increase the sampling rate to q=0.5q=0.5 to reduce the impact of DP noise, while keeping the number of training epochs fixed at 150, resulting in 300 updates. We recalibrate the noise multiplier to maintain the same privacy budget.

To evaluate this choice, we compare the adapted baseline schedule with DPLongSyn's schedule on ADNI at ε=8\varepsilon=8, recalibrating the noise multiplier for each schedule to maintain the same privacy budget.

MethodFID ↓\downarrow (q=0.5q=0.5, 300 updates)FID ↓\downarrow (q=0.05q=0.05, 5,540 updates)
DP-Indep137.7322.8
DP-TabImage164.1286.1

Both baselines achieve substantially better FID under the adapted schedule, supporting our choice of training configuration.

Checkpoint selection. For public pretraining, we use fixed training horizons of 150 epochs for DP-Indep, DP-TabImage, DP-ImageTab, and DP-JointDM, and 50 epochs for DP-JointMERF. All methods use the final public checkpoint to initialize subsequent training. Similarly, for private DP-SGD training, we use the final checkpoint after 300 updates.

We have included these training and checkpoint selection details in Appendix B.5 in the updated version.

Q4: Explanation of high FID values.

The relatively high FID values are largely due to the domain gap between public source images and real follow-up images. As shown in Table 29 (Matched transition controls), even when we directly use real public images as the generated follow-ups without any deformation, the FID against real private follow-up images remains high (88.9 on ADNI and 132.3 on OASIS-3). This suggests that the high FID reflects a distribution mismatch between public and private images. We have also visually inspected the generated images and found their visual quality to be reasonable, although we could not include examples in the paper due to dataset usage restrictions.


We thank the reviewer again for the helpful comments. If our responses have addressed your concerns, we would appreciate it if you could consider raising your rating. We would also be happy to address any further questions you may have.

AuthorsOpenReview ↗

A gentle reminder

Dear Reviewer ctS7,

We've taken your feedback into careful consideration in our response. Could you please check whether our response has properly addressed your concern? If so, could you please kindly consider increasing your initial score accordingly? Certainly, we are more than happy to answer your further questions.

Thank you for your time and effort in reviewing our work!

Best Regards,

Authors

Thank you for the clarifications and additional experiments. I’m keeping my original rating.

Thank you very much for your feedback and for maintaining your positive rating. We really appreciate your time and effort!

Reviewer_ApYNOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper works on differentially private (DP) synthesis of longitudinal image data. The proposed method works in three stages: 1) Train a registration model and a transition model on public longitudinal data; 2) Generate synthetic records and privately fine-tune the transition model with DP; 3) At usage, take a public source image, sample a deformation and warp the source to produce the follow-up image. Experiments are conducted on three datasets.

Critical strengths
  1. The framework in Figure 1 is interesting and reasonable. Public data provides the appearance and transition prior, while the private data is used only to learn how transitions depend on private data.

  2. Overall, the experimental results are good. The method improved over compared methods on multiple datasets and various settings.

Critical weaknesses
  1. There is no non-DP reference for image quality. The paper does not show how much quality is lost to privacy.

  2. It would be interesting to see an ablation study on how many public images are used for training, to investigate how much the method relies on public data and to better understand the contributions of public data and private data.

  3. The method generates a deformation through a registration model. An obvious alternative is to train a conditional generator of the follow-up image directly from the source image. An ablation study is missing to show whether the registration model is needed.

  4. The paper does not discuss/compare with DP generation of videos or other sequential data, which is closely related: longitudinal pairs are essentially short sequences. The claim of being "the first" is an overclaim.

  5. How does the domain gap between the public data and the private data affect the model? What if the domain gap is large?

Questions for authors

See above.

4 comments
AuthorsOpenReview ↗

Responses to Reviewer ApYN (Part 1)

We thank the reviewer for the careful reading and thoughtful comments. We address the reviewer's questions in the following and have revised the paper accordingly. The changes are marked in blue in our revision. We hope the responses below address the reviewer's concerns.

W1: There is no non-DP reference for image quality. The paper does not show how much quality is lost to privacy.

We thank the reviewer for the suggestion. To quantify the impact of privacy on image quality, we additionally evaluate a non-private version of DPLongSyn (ε=∞\varepsilon=\infty), using the same training procedure but without adding DP noise. We report the FID scores and compare them with those obtained under different privacy budgets.

Datasetε=∞\varepsilon=\infty (non-DP)ε=32\varepsilon=32ε=8\varepsilon=8ε=2\varepsilon=2
ADNI92.393.893.894.3
OASIS-3154.1152.3153.9146.5
CelebV-HQ112.8113.5116.1116.7

The results show that DPLongSyn maintains image quality comparable to the non-private reference, demonstrating a favorable privacy–utility trade-off.

We have included this comparison in Appendix B.10 of the updated paper.

W2: How sensitive is DPLongSyn to the amount of public pretraining data, and what does private fine-tuning contribute?

We vary the number of public DLBS image pairs used to pretrain the transition model and report results on ADNI at ε=8\varepsilon=8.

Public image pairsFID ↓\downarrowRecord–image change correlation ρ\rho ↑\uparrow
13493.40.132
26794.30.182
53593.80.161

The results show that DPLongSyn maintains similar FID and record–image change correlation even when the public pretraining data is reduced from 535 to 134 pairs, suggesting that it can perform well with relatively limited public data. We have included this ablation in Appendix B.11 of the updated paper.

To isolate the contribution of private fine-tuning, we compare follow-up images generated by the publicly pretrained transition model and the privately fine-tuned transition model.

Follow-up generationFID ↓\downarrowRecord–image change correlation ρ\rho ↑\uparrow
Publicly pretrained transition model88.6-0.075
Privately fine-tuned transition model93.8+0.161

The results show that private fine-tuning is important for learning meaningful longitudinal changes. Although the publicly pretrained transition model achieves a slightly better FID, its generated changes have a negative correlation (-0.075) with the changes described in the records. After private fine-tuning, this correlation increases to +0.161, indicating better alignment between the generated anatomical changes and the corresponding records.

W3: How does the registration-based model compare with a conditional generator that predicts the follow-up image directly from the source image?

We conduct an ablation study on ADNI at ε=8\varepsilon=8, comparing two approaches: (1) predicting a deformation field to warp the source image into the follow-up image, and (2) directly predicting the follow-up image using a conditional generator.

The registration-based approach achieves better record–image change correlation ρ\rho (0.161 vs. 0.125), indicating that the generated longitudinal changes better align with the corresponding records.

We have included this comparison in Appendix B.12 of the updated paper.

AuthorsOpenReview ↗

Responses to Reviewer ApYN (Part 2)

W4: How does DPLongSyn relate to prior work on DP video and sequence synthesis?

We thank the reviewer for highlighting this connection. We have expanded the Related Work section to discuss this connection.

Prior work has studied DP for video release and sequence generation. VideoDP (Wang et al., 2020) produces a sanitized version of an input video for downstream analysis, while our goal is to generate synthetic longitudinal images. Esteban et al. (2017) use a DP recurrent GAN to generate medical time series, and Matsumoto et al. (2024) use a DP diffusion model to generate sequential signals. These methods focus on numerical time-series data, while our setting involves longitudinal image pairs and associated records describing the visual changes between observations.

DPLongSyn is the first framework to synthesize longitudinal images together with records describing changes across time points under differential privacy. We have revised the abstract to make the scope of our claim more precise.

Wang, Han, Shangyu Xie, and Yuan Hong. "Videodp: A flexible platform for video analytics with differential privacy." Proceedings on Privacy Enhancing Technologies (2020).

Esteban, Cristóbal, Stephanie L. Hyland, and Gunnar Rätsch. "Real-valued (medical) time series generation with recurrent conditional gans." arXiv preprint arXiv:1706.02633 (2017).

Matsumoto, Tomoya, et al. "Differentially Private Sequential Data Synthesis with Structured State Space Models and Diffusion Models." Neurips Safe Generative AI Workshop 2024. 2024.

W5: How does the domain gap between the public data and the private data affect the model? What if the domain gap is large?

We would like to highlight that all three experimental settings involve a public–private domain gap. For brain MRI, the public DLBS cohort contains only cognitively normal adults, while the private ADNI and OASIS-3 cohorts include participants with cognitive impairment. For facial expressions, public KDEF images are captured under controlled conditions, while private CelebV-HQ frames come from Internet videos.

DPLongSyn addresses these differences through public source selection and private fine-tuning. For brain MRI, we use age and diagnosis information from synthetic records to select public source images with corresponding anatomical characteristics. To evaluate its effectiveness, we compare mean source hippocampal volumes across diagnosis groups on ADNI at ε=2\varepsilon=2.

Mean hippocampal volume (mm³)CNMCIDementia
ADNI source images7,3276,7635,703
DLBS source images with random selection8,4438,4588,497
DLBS source images with age and diagnosis matching8,6038,0197,205

The results show that our source selection better captures anatomical differences across diagnosis groups. For example, the CN-to-dementia difference in hippocampal volume is 1,398 mm³, closer to the 1,624 mm³ observed in real ADNI, compared with the -54 mm³ obtained with random selection.

Private fine-tuning further helps bridge the domain gap by learning how longitudinal changes relate to the corresponding records in the target population. As shown in our response to W2, private fine-tuning improves record–image change correlation compared with the publicly pretrained transition model.

We have included this discussion in Appendix B.13 of the updated paper.


We thank the reviewer again for the helpful comments. If our responses have addressed your concerns, we would appreciate it if you could consider raising your rating. We would also be happy to address any further questions you may have.

Thanks, I have increased the score.

Thank you very much for your positive feedback! Your valuable suggestions have helped us improve the quality of our work. We really appreciate it.

Reviewer_mYBvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes DPLongSyn which is a framework that generates image–record–image tuples under differential privacy. It first learns a deformation representation and conditional diffusion model from public image pairs. It then privately synthesizes structured records and adapts selected diffusion-model parameters using DP-SGD. During generation, it selects a public source image and applies a record-conditioned deformation to produce the follow-up. The paper provides theoretical and experimental evidence.

Critical strengths
  1. The paper addresses an important problem. Using public source images while privately learning record-conditioned transitions exploits the shared visual content between longitudinal observations. This provides a useful approach to reducing the private learning burden, and it is a significant contribution to private synthetic-data generation.

  2. The experimental evaluation is rigorous. The method is evaluated on image fidelity and classification utility using three different privacy budgets and three datasets.

Critical weaknesses
  1. The baselines seem to generate observations, but DPLongSyn uses a real public source. That might be a design advantage, but it complicates attribution. Furthermore, the training protocols also differ: DPLongSyn uses sampling rates around 0.05 and over 5,000 steps (Table 7), while the adapted baselines use a sampling rate of 0.5 and 300 steps (Appendix B5). These differences might not be unfair, but their tuning needs stronger justification.

  2. The generated distribution uses qpub(x0∣a)q_{\mathrm{pub}}(x_0\mid a) for baseline appearance and not the learning the private baseline distribution. Thus, private baseline characteristics absent from the public pool cannot generally be recovered through source selection. This seems important because the DLBS dataset contains cognitively normal adults, while the private MRI cohorts include cognitive impairment. The age-adjusted atrophy ranking used for source selection may help, but its exact procedure and contribution need clearer evaluation. This matching is not itself privacy leakage when it uses only public data and DP records. The concern is fidelity and whether source matching accounts for much of the observed classification performance.

Questions for authors
  1. Do private transitions improve progression prediction over the matched public-transition and zero-deformation controls?

  2. I was wondering if the improvements stable across independent training/synthesis seeds and paired confidence intervals?

  3. How much utility comes from diagnosis-conditioned public-source matching?

  4. Could you please report what data are used for latent-scale calibration?

  5. I was wondering if the face splits are identity-disjoint, and what privacy unit is claimed there?

4 comments
AuthorsOpenReview ↗

Responses to Reviewer mYBv (Part 1)

We thank the reviewer for the careful reading and thoughtful comments. We address the reviewer's questions in the following and have revised the paper accordingly. The changes are marked in blue in our revision. We hope the responses below address the reviewer's concerns.

W1: How much of DPLongSyn’s advantage comes from using a public source, and do the different DP-SGD schedules affect the comparison?

Contribution of the public source. DPLongSyn and all baselines use the same public dataset for pretraining. The key difference is that DPLongSyn additionally uses a public image as the source image during synthesis. This design takes advantage of a key property of longitudinal data: much of a subject’s anatomy remains unchanged across visits, while only certain features evolve over time. Therefore, instead of generating the entire follow-up image from scratch, DPLongSyn uses a public image to provide the underlying anatomy and learns a private transition model to capture how it changes over time.

To separate the benefit of using a public source from that of private fine-tuning, we conduct an ablation study on ADNI under ε=8\varepsilon=8. We keep the public source images and records fixed and compare three ways of generating the follow-up image: (1) directly copying the public source image, (2) applying the publicly pretrained transition model, and (3) applying the privately fine-tuned transition model.

Follow-up generationFID ↓\downarrowRecord–image change correlation ρ\rho ↑\uparrow
Copy the public image88.9N/A (zero image change)
Apply the publicly pretrained transition model88.6-0.075
Apply the privately fine-tuned transition model93.8+0.161

The results show two different contributions. First, the public source explains the strong FID performance. Simply copying the public image achieves an FID of 88.9, already much better than the strongest baseline (137.7). This indicates that the FID advantage mainly comes from leveraging existing anatomical information.

Second, private fine-tuning is important for learning meaningful longitudinal changes. Although the publicly pretrained transition model achieves a slightly better FID, its generated changes have a negative correlation (-0.075) with the changes described in the records. After private fine-tuning, this correlation increases to +0.161, indicating better alignment between the generated anatomical changes and the corresponding records.

Effect of the training schedule. DPLongSyn uses a fixed training schedule without tuning on private data. For the baselines, we follow the original papers' 150 training epochs. However, using a sampling rate of q=0.05q=0.05 on our smaller private datasets results in small batches and noisy updates. We therefore increase the sampling rate to q=0.5q=0.5 to reduce the impact of DP noise, while keeping the number of training epochs fixed at 150, resulting in 300 updates. We recalibrate the noise multiplier to maintain the same privacy budget. Our results show that both baselines achieve better FID under the adapted schedule, supporting this choice.

To study whether differences in DP-SGD schedules affect the comparison, we evaluate DPLongSyn and the baselines under two schedules: (1) q=0.5q=0.5 with 300 updates, and (2) q=0.05q=0.05 with 5,540 updates. For each schedule, we recalibrate the noise multiplier to maintain the same privacy budget (ε=8\varepsilon=8).

MethodFID ↓\downarrow at q=0.5q=0.5, 300 updatesFID ↓\downarrow at q=0.05q=0.05, 5,540 updates
DP-Indep137.7322.8
DP-TabImage164.1286.1
DPLongSyn94.993.8

Both baselines achieve better FID under the adapted schedule, supporting our choice of training configuration. Additionally, DPLongSyn outperforms baselines under the same training schedule.

We have included the additional experiments in Appendix B.5 of the updated paper.

AuthorsOpenReview ↗

Responses to Reviewer mYBv (Part 2)

W2: How well does public-source matching reproduce private source anatomy, and how much does it contribute to follow-up diagnosis classification?

Source selection. DPLongSyn selects source images from the public DLBS dataset, which contains only cognitively normal participants. For each synthetic record, we first identify the 40 DLBS images closest in age and rank them by an atrophy score based on hippocampal and ventricular volumes. We then sample an image from a rank interval determined by the source and follow-up diagnoses in the synthetic record. For example, CN-to-CN records select from the lowest 40% of atrophy scores, while dementia-to-dementia records select from the highest 20%.

Source fidelity. To evaluate how well public-source matching captures diagnosis-related anatomical differences, we compare mean source hippocampal volumes across diagnosis groups on ADNI at ε=2\varepsilon=2.

Mean hippocampal volume (mm³)CNMCIDementia
ADNI source images7,3276,7635,703
DLBS source images with random selection8,4438,4588,497
DLBS source images with age matching8,1278,2018,149
DLBS source images with age and diagnosis matching8,6038,0197,205

Our results show that random selection and age matching alone produce similar hippocampal volumes across diagnosis groups, while diagnosis-conditioned matching captures the anatomical differences between groups. For example, the CN-to-dementia difference in hippocampal volume is 1,398 mm³, closer to the 1,624 mm³ observed in real ADNI, compared with the -54 mm³ obtained with random selection.

Classification utility. To quantify how source matching contributes to downstream classification, we compare age-only and diagnosis-conditioned matching under two settings: (1) directly copying the selected source image as the follow-up, and (2) generating the follow-up using the privately trained transition model. We report the follow-up diagnosis classification AUC on ADNI at ε=2\varepsilon=2.

Follow-up generationAge-only matchingAge + diagnosis matching
Copy source image0.5020.568
Private transition model0.4970.577

The results show that our public source matching step improves classification performance on ADNI. Even without applying the transition model, copying the matched source image achieves an AUC of 0.568, which further increases to 0.577 with the privately trained transition model. This is expected, as source selection uses source and follow-up diagnoses to select images with corresponding anatomical characteristics.

We would also like to highlight that DPLongSyn improves downstream classification on CelebV-HQ, where no source matching is used. In addition, private transition training improves the consistency between generated longitudinal changes and the corresponding records. These results show that source matching is not the only factor contributing to DPLongSyn's performance.

We have included this discussion in Appendix B.13 of the updated paper.

Q1: Do private transitions improve progression prediction over the matched public-transition and zero-deformation controls?

To evaluate the contribution of private transition learning to progression prediction, we conduct an ablation study on ADNI at ε=2\varepsilon=2, varying how the follow-up images are generated: (1) copying the source image (zero deformation), (2) applying the publicly pretrained transition model, and (3) applying the privately fine-tuned transition model.

How the follow-up is generatedProgression AUC
Copy the source image (zero deformation)0.555
Apply the publicly pretrained transition model0.549
Apply the privately fine-tuned transition model0.579

The results show that private transition learning improves progression prediction, achieving an AUC of 0.579 compared with 0.555 for zero deformation and 0.549 for the public transition model.

We have included these results in Appendix B.9 of the updated paper.

AuthorsOpenReview ↗

Responses to Reviewer mYBv (Part 3)

Q2: I was wondering if the improvements stable across independent training/synthesis seeds and paired confidence intervals?

To evaluate the stability of our results, we repeat the experiments across 5 independent training and synthesis seeds on ADNI at ε=8\varepsilon=8. We report the mean performance and standard deviation in the table below.

FID ↓\downarrowKID ↓\downarrowTVD ↓\downarrowρ\rho ↑\uparrowFollow-up diagnosis AUC ↑\uparrowProgression AUC ↑\uparrow
Best baseline138.96±8.16138.96 \pm 8.160.1804±0.00670.1804 \pm 0.00670.0506±0.00370.0506 \pm 0.00370.0260±0.00660.0260 \pm 0.00660.5228±0.01260.5228 \pm 0.01260.5374±0.02110.5374 \pm 0.0211
DPLongSyn93.56±0.8693.56 \pm 0.860.0974±0.00270.0974 \pm 0.00270.0318±0.00450.0318 \pm 0.00450.1758±0.01800.1758 \pm 0.01800.5524±0.00610.5524 \pm 0.00610.5636±0.02920.5636 \pm 0.0292

We present the paired 95% confidence intervals for the improvements over the strongest baseline in the following.

MetricMean improvementPaired 95% CI
FID ↓\downarrow45.40[34.52, 56.28]
KID ↓\downarrow0.0830[0.0724, 0.0936]
TVD ↓\downarrow0.0188[0.0110, 0.0266]
ρ\rho ↑\uparrow0.1498[0.1323, 0.1673]
Follow-up diagnosis AUC ↑\uparrow0.0296[0.0150, 0.0442]
Progression AUC ↑\uparrow0.0262[-0.0206, 0.0730]

The results show that DPLongSyn consistently achieves better mean performance across all six metrics.

We have included these additional results in Appendix B.14 of the updated paper.

Q3: How much utility comes from diagnosis-conditioned public-source matching?

To study the contribution of source selection, we compare three matching rules on ADNI under ε=8\varepsilon=8: (1) random selection (no matching), (2) age matching, and (3) age and diagnosis matching.

Public-source selectionFID ↓\downarrowKID ↓\downarrowρ\rho ↑\uparrowFollow-up diagnosis AUC ↑\uparrowProgression AUC ↑\uparrow
Random96.50.1030.1120.5290.524
Age only93.70.0970.1370.5020.525
Age + diagnosis93.80.0990.1610.5610.559

The results show that age matching improves image quality, while adding diagnosis-conditioned matching further improves longitudinal consistency and downstream classification performance.

We have included these results in Appendix B.13 of the updated paper.

Q4: Could you please report what data are used for latent-scale calibration?

For ADNI and OASIS-3, we calibrate the latent scale on public DLBS validation pairs. We use the DP synthetic records to estimate the target deformation magnitude relative to DLBS, then refine the scale using 160 pilot pairs generated from those records and public DLBS source images. For CelebV-HQ, we use public KDEF validation pairs and select the scale that best matches the spread of deformations measured on those pairs. In all cases, calibration uses only public data and outputs of the DP mechanisms, so it requires no additional privacy budget.

We have clarified the calibration data and procedure in Appendix B.4 of the updated paper.

Q5: I was wondering if the face splits are identity-disjoint, and what privacy unit is claimed there?

In CelebV-HQ, one original YouTube video may provide several short facial-expression segments. From each segment, we construct one longitudinal sample consisting of a neutral frame, a frame near peak expression, and a record describing the change. For example, our 779 training samples come from 613 original videos. Since CelebV-HQ does not provide person IDs, we cannot verify whether the same person appears in different videos.

In DPLongSyn and all baselines, one privacy unit consists of both frames and the associated record from one segment. Our framework can also be extended to video-level DP by treating all samples from the same video as a single privacy unit and adjusting the privacy accounting accordingly. Individual-level DP would additionally require identity information to group samples belonging to the same individual.


We thank the reviewer again for the helpful comments. If our responses have addressed your concerns, we would appreciate it if you could consider raising your rating. We would also be happy to address any further questions you may have.

AuthorsOpenReview ↗

A gentle reminder

Dear Reviewer mYBv,

We've taken your feedback into careful consideration in our response. Could you please check whether our response has properly addressed your concern? If so, could you please kindly consider increasing your initial score accordingly? Certainly, we are more than happy to answer your further questions.

Thank you for your time and effort in reviewing our work!

Best Regards,

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.