Two papers have been accepted to USENIX Security 2026!

Created on December 20, 2025

2025

Window-Based Comparison (WBC)
Imitative Membership Inference Attack (IMIA)

I am thrilled to share that two of my worked papers have been accepted to USENIX Security 2026! Both works focus on advancing the understanding of privacy risks in machine learning models through novel Membership Inference Attacks (MIAs).

1. Window-based Membership Inference Attacks Against Fine-tuned LLMs

The first paper, Window-based Membership Inference Attacks Against Fine-tuned Large Language Models, challenges the prevailing global-averaging paradigm used in existing MIAs.

In this work, we discovered that membership signals in fine-tuned LLMs often manifest as sparse, extremal events that are easily diluted by global statistics. To address this, we introduced Window-Based Comparison (WBC), a method that slides windows of varying sizes across text sequences to capture these localized memorization patterns. Our experiments demonstrate that WBC substantially outperforms established baselines, achieving 2-3x improvements in detection rates at low false positive thresholds. For a deeper dive into the intuition behind WBC, check out the blog post!

2. Imitative Membership Inference Attack

The second paper, Imitative Membership Inference Attack, addresses the prohibitive computational costs associated with state-of-the-art shadow training methods.

We propose IMIA, which employs a novel imitative training technique to strategically construct a small number of target-informed models. These models closely replicate the target model’s behavior, allowing for highly effective inference without the massive overhead of training hundreds of shadow models. IMIA achieves state-of-the-art performance while requiring less than 5% of the computational cost of existing approaches.

For a deeper dive into the intuition behind this work, check out Yuntao’s awesome blog post!

A huge thank you to my collaborators Yuntao Du and Kaiyuan Zhang, and to Prof. Bruno Ribeiro and Prof. Hanshen Xiao for their expert advice. I also gratefully acknowledge the support from Ashish Kundu and Charles Fleming at Cisco. Finally, special thanks to my advisor Prof. Ninghui Li for his invaluable guidance.

I look forward to presenting these works in August! Boiler Up! 🚂