Auditing a Conditional-VAE Synthetic-Data Firewall for Privacy-Preserving Deep Learning

Privacy-sensitive machine learning programmes need deployable models without repeatedly circulating original records among developers, analysts and downstream services. This paper evaluates a class-conditional variational autoencoder (CVAE) placed as a synthetic-data firewall between a protected corpus and downstream learners that never see the original records, and contributes an audit and release protocol for such a pipeline. The protocol separates four adversary tiers (model, release, sampling and white-box generator access), calibrates membership attacks with 32 shadow pipelines, adds deliberately vulnerable positive controls, tests copying in the lower tail of the distance distribution, and gates the release on utility, exposure, copying and worst-class retention with an exposure threshold anchored to chance. On the digit corpus, over ten independent partitions, synthetic-only training retains % (CNN) and % (GRU) of real-data accuracy. On digits and a clinical table, likelihood-ratio attacks detect leakage in real-data models but not in release-trained models, and memorising learners are protected on all three corpora (including a face-image set). The firewall relocates rather than removes risk: adversaries holding the release or the generator retain a measurable membership signal, and with very small cohorts, the generator passes it through. Conditional GAN, Gaussian-mixture and differentially private baselines, alternative likelihoods, and an independent selection protocol (% for the selected configuration) complete the evaluation; empirical resistance is not a formal (ε,δ) guarantee.

Authors

Institutions

Publication Details

Journal
Future Internet
Published
2026-11-16
DOI
https://doi.org/10.3390/fi18110632
Primary Topic
Privacy-Preserving Technologies in Data
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Auditing a Conditional-VAE Synthetic-Data Firewall for Privacy-Preserving Deep Learning

Sami Aziz Alshammari, Osama Harfoushi, Mamoon Obiedat, Ayoub Alsarhan et al.
Future Internet
Privacy-Preserving Technologies in Data
article

Auditing a Conditional-VAE Synthetic-Data Firewall for Privacy-Preserving Deep Learning

Sami Aziz Alshammari, Osama Harfoushi, Mamoon Obiedat, Ayoub Alsarhan, Tamam Alsarhan, Malek Mahmoud Barhoush, Mohammad Khasawneh
article en

Abstract

Privacy-sensitive machine learning programmes need deployable models without repeatedly circulating original records among developers, analysts and downstream services. This paper evaluates a class-conditional variational autoencoder (CVAE) placed as a synthetic-data firewall between a protected corpus and downstream learners that never see the original records, and contributes an audit and release protocol for such a pipeline. The protocol separates four adversary tiers (model, release, sampling and white-box generator access), calibrates membership attacks with 32 shadow pipelines, adds deliberately vulnerable positive controls, tests copying in the lower tail of the distance distribution, and gates the release on utility, exposure, copying and worst-class retention with an exposure threshold anchored to chance. On the digit corpus, over ten independent partitions, synthetic-only training retains % (CNN) and % (GRU) of real-data accuracy. On digits and a clinical table, likelihood-ratio attacks detect leakage in real-data models but not in release-trained models, and memorising learners are protected on all three corpora (including a face-image set). The firewall relocates rather than removes risk: adversaries holding the release or the generator retain a measurable membership signal, and with very small cohorts, the generator passes it through. Conditional GAN, Gaussian-mixture and differentially private baselines, alternative likelihoods, and an independent selection protocol (% for the selected configuration) complete the evaluation; empirical resistance is not a formal (ε,δ) guarantee.

Future InternetVol. 18(11)
Al-Ahliyya Amman University (JO), University of Jordan (JO), Northern Border University (SA), Higher Colleges of Technology (AE), Hashemite University (JO), Jadara University (JO), Yarmouk University (JO)
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon