A membership inference attack against Private Evolution
Start here
An arXiv paper proposes a white-box membership inference attack against Private Evolution (PE), which the authors state is the first such attack targeting PE's underlying voting algorithm. PE is a class of inference-only methods that generate synthetic data under differential privacy. The attack exploits the property that iterative filtering biases synthetic candidates toward private instances, and achieves empirical results showing private data leakage in PE applications across image, text, and tabular modalities.
Generated by AI from reports · Updated 21 days ago
Attention and developments
Attention, hourly
Now 30, 0% in 24 hours·Peak 30 (Nov 6 04:00)
Related research
This event's paper
- Membership Inference Attack on Private Evolution
arXiv · 2026-10-21 · Deep reading
Progress timeline
- arXiv: Adversarial Attacks, Backdoors & PoisoningMembership Inference Attacks against Private Evolution
The researchers propose a white-box membership inference attack that the authors describe as the first such attack against the underlying voting algorithm of the Private Evolution (PE) framework. PE is a class of inference-only synthetic data generation methods that generate high-quality data with black-box access to pretrained foundation models, and is designed to produce synthetic data under differential privacy (DP). The attack exploits the property that the iterative filtering process biases the survival and proximity of synthetic candidates toward private instances, and obtains empirical results indicating private data leakage in PE applications across image, text, and tabular modalities. The paper also notes that formal DP guarantees do not exempt released synthetic data from empirical privacy audits.