Medical model training changed privacy risk more than who ranked highest
Source check: Source checked. Not independent scientific confirmation.
Finding: Across six clinical datasets, privacy-risk rankings were broadly similar under many changes to model training. In one chest X-ray cohort, stronger regularization reduced the share of patients above the study’s risk threshold from 88% to 30%. The rankings were compared across training changes using an audit of 7,789 medical-model instances.
Technology: ResNet, EfficientNet, vision transformers, ECGFounder, MRI-CORE
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Researchers trained several thousand medical-model instances, including ECGFounder and MRI-CORE adaptations, and compared aggregate distributions of measured privacy risk across training regimes.
- Test and result
- A defensive study compared aggregate risk distributions and ordinal stability under seed, architecture, regularization and fine-tuning changes, including comparisons with trained seed replicates.
- Limitation / still unknown
- The outcomes measure estimated likelihood of training-set membership disclosure under the study’s audit. They do not represent observed patient disclosures or complete invariance across all settings.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.30229 · https://arxiv.org/pdf/2610.30229v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: arXiv v1 PDF affiliation, aggregate Figures 2 and 4, cohort and risk-stability comparison.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.