The pretraining domain outweighs the training objective in setting the privacy-utility trade-off of differentially private medical image analysis

2026-01-27 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition · Depth: Expert, quick

Summary

A study on differentially private medical image analysis demonstrates that the pretraining domain significantly outweighs the training objective in determining the privacy-utility trade-off. Researchers trained ConvNeXt classifiers using differentially private stochastic gradient descent from five initializations, varying both objective and domain independently. Evaluated on over 590,000 chest radiographs from five external datasets, supervised pretraining on chest radiographs consistently ranked first in 24 of 25 dataset and budget combinations. Its lead over ImageNet pretraining grew from 2.5 to 14.6 points of macro-averaged area under the receiver operating characteristic curve as privacy budgets tightened, with the domain effect exceeding the objective effect by a factor of 2.2 to 3.4. Pretraining the corpus privately cost about 5 points but still outperformed every public initialization under privacy. Low-Rank Adaptation (LoRA) removed about half the residual gap, and in-domain pretraining improved the worst-performing demographic subgroup.

Key takeaway

For AI Scientists developing differentially private medical imaging models, your strategy for pretraining data is paramount. Prioritize supervised pretraining on in-domain medical datasets, such as chest radiographs, even if it requires private pretraining. This approach significantly improves diagnostic accuracy under strict privacy budgets and enhances fairness for demographic subgroups, outperforming models pretrained on generic public datasets like ImageNet. Consider integrating Low-Rank Adaptation (LoRA) to further optimize utility.

Key insights

Pretraining domain is more critical than the training objective for utility in differentially private medical imaging.

Principles

• Domain-specific pretraining enhances utility under differential privacy.
• Private pretraining on relevant data can outperform public, generic pretraining.

Method

ConvNeXt classifiers were trained with differentially private stochastic gradient descent from five initializations, varying pretraining objective and domain, then evaluated on 590,000 chest radiographs.

In practice

• Prioritize supervised pretraining on in-domain medical images.
• Apply Low-Rank Adaptation (LoRA) to mitigate utility gaps.

Topics

• Differential Privacy
• Medical Imaging
• Pretraining Domain
• ConvNeXt
• Chest Radiographs
• Low-Rank Adaptation

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, AI Security Engineer

Related on AIssential

• A New Privacy-First AI Predicts COVID Severity Using X-Rays and Medical Records — A privacy-first AI predicts COVID-119 severity using federated learning on X-rays and medical records.
• Contrast-Informed Augmentation and Domain-Adversarial Training for Adult-to-Neonatal MR Reconstruction Generalization — Contrast-informed augmentation and domain-adversarial training enhance deep learning MR reconstruction generalization from adult to neonatal data.
• OpenMedQ: Broad Open Pretraining for Medical Vision-Language Models — Broad open pretraining on diverse medical data enables SOTA performance with smaller models.
• Self-supervision drives representational convergence in medical foundation models more than clinical supervision — Self-supervised objectives, not scale or clinical supervision, primarily drive representational convergence in medical foundation models.
• Demographically-Conditioned Synthetic Medical Images for Bias Mitigation and Bias Detection in Disease Classifiers — Demographically-conditioned synthetic medical images effectively mitigate and detect bias in disease classifiers, especially for minority subgroups.
• IEEE Transactions on Artificial Intelligence, Volume 7, Issue 6, June 2026 — AI research in June 2026 emphasizes robustness, privacy, efficiency, and specialized applications across diverse domains.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.