Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

Membership Inference or Data Split Bias? Identifying False Positives in Synthetic Medical Image Privacy Audits.

Linxue Bai1, Omid Pournik1, Xuefei Ding1

  • 1University of Birmingham, UK.

Studies in Health Technology and Informatics
|May 23, 2026
PubMed
Summary
This summary is machine-generated.

Related Concept Videos

Bias in Epidemiological Studies01:29

Bias in Epidemiological Studies

Biases can arise at various stages of research, from study design and data collection to analysis and interpretation. Recognizing and addressing these biases is essential to ensure the validity and reliability of epidemiological findings.Broadly speaking, biases in epidemiology fall into three main categories: selection bias, information bias, and confounding. A more detailed description of possible biases is:

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Association of maternal serum copper during early pregnancy with the risk of spontaneous preterm birth: A nested case-control study in China.

Environment international·2018
Same author

Automatic Assessment of Full Left Ventricular Coverage in Cardiac Cine Magnetic Resonance Imaging with Fisher Discriminative 3D CNN.

IEEE transactions on bio-medical engineering·2018
Same author

Diphenyl Ethers from a Marine-Derived <i>Aspergillus sydowii</i>.

Marine drugs·2018
Same author

In situ determination of trace elements in melt inclusions using laser ablation inductively coupled plasma sector field mass spectrometry.

Rapid communications in mass spectrometry : RCM·2018
Same author

A genetic variant in LINGO2 contributes to the risk of gestational diabetes mellitus in a Chinese population.

Journal of cellular physiology·2018
Same author

Gap Activity TOward Rags 1 variants in Chinese people with sporadic drug-resistant focal epilepsy.

Acta neurologica Scandinavica·2018

Membership Inference Attacks (MIAs) may falsely flag synthetic medical data privacy leaks. These attacks can detect experimental bias instead of true membership signals, necessitating control experiments for reliable privacy auditing.

Area of Science:

  • Medical Imaging
  • Artificial Intelligence
  • Data Privacy

Background:

  • Membership Inference Attacks (MIAs) are standard for auditing synthetic medical data privacy.
  • High MIA scores are typically interpreted as significant privacy leaks.

Purpose of the Study:

  • To investigate if MIAs detect true membership signals or experimental artifacts like data split bias.
  • To assess the reliability of MIA results in synthetic medical data privacy audits.

Main Methods:

  • Experiments conducted on synthetic brain MRI images generated by a StyleGAN3 model.
  • A Primary MIA (member vs. non-member data) was designed.
  • A Control Experiment using only member data subsets was implemented to isolate data split bias.
Keywords:
Data Split BiasFalse PositivesGenerative ModelsMembership Inference Attack (MIA)Privacy AuditingSynthetic Medical Images

Related Experiment Videos

Main Results:

  • The Control Experiment, lacking true membership distinction, achieved AUCs matching or exceeding the Primary MIA.
  • AUC improvements were +0.0205 on the 5k-pool Small corpus and +0.0113 on the 30k-pool Big corpus.
  • This indicates classifiers may learn statistical artifacts rather than genuine privacy risks.

Conclusions:

  • High MIA scores can be false positives for privacy breaches.
  • Rigorous control experiments are essential to differentiate genuine privacy risks from experimental bias.
  • MIA results are unreliable for synthetic medical data privacy auditing without validation against bias.