[论文] OP-Print: Tracing Membership Imprints in On-Policy Self-Distillation

论文概要

研究领域: ML
作者: Qi Li, Xinchao Wang
发布时间: 2026-10-22
arXiv: 2610.31616

中文摘要

在策略自蒸馏通过使用学生自身输出作为监督来消除外部教师需求，但其隐私风险仍未被充分探索——同一训练样本的两个耦合视图可能加强记忆。本文首次提出对在策略自蒸馏的成员推理攻击研究，提出统一攻击框架OP-Print。核心发现：自蒸馏中的监督印记是稀疏且有结构的，集中在不确定性敏感的纠正位置。灰盒设置中直接聚合教师-学生残差纠正信号；黑盒设置中以重启输出盆地稳定性作为替代指标。所有基线攻击即使在灰盒访问下也仅相当于随机猜测，而OP-Print在两种设置中均被广泛验证有效。

原文摘要

On-policy self-distillation removes the need for an external teacher by using the student's own outputs as supervision, yet its privacy risks remain largely unexplored. Using two coupled views of the same training example may strengthen its memorization. We present the first study of membership inference attacks against on-policy self-distillation, proposing OP-Print, a unified attack framework. OP-Print builds on our finding that the supervision imprint in on-policy self-distillation is sparse and structured, concentrating on uncertainty-sensitive correction sites. In gray-box settings, OP-Print aggregates the privileged teacher-student residual correction along the same rollout; in black-box settings, it measures restart-based output-basin stability as a surrogate. All baseline attacks r...

自动采集于 2026-10-26

#论文 #arXiv #ML #小凯