EVENT DOSSIER
On September 7, 2026, a study published on arXiv cs.LG indicated that there is a risk of privacy leakage in the federal forgetting system. Malicious clients can use the linear classifiers broadcast by the server to derive the hidden abstract features of deleted samples through reverse analysis. Experiments based on the MNIST and CIFAR-10 datasets showed that when the broadcast accuracy was high, attackers could accurately recover the deletion labels for each test sample; while low-precision broadcasts reduced the ability to recover detailed information, but did not sufficiently improve response diversity or even completely block recognition. The study further characterized the conditions under which the observed data contains sufficient independent information, proposed an optimal construction method for unrestricted detection, and developed a realistic estimator based on the attacker’s own data. It also analyzed the actual impact of broadcast accuracy, update verification, response rate, and concurrent activities on system privacy and integrity.
本文指出，在联邦遗忘系统中，恶意客户端可利用广播的线性分类器逆向推导出被删除样本的隐藏摘要。研究者在 MNIST 和 CIFAR-10 数据集上实验表明，高精度广播允许对每个测试样本的删除进行精确标签恢复；而低精度广播则显著降低细粒度恢复能力，响应多样性不足甚至完全阻止识别。文章刻画了观测数据包含足够独立信息的条件，给出了无限制探测的最优构造及基于攻击者自身数据的更现实估计器，并分析了广播精度、更新验证、响应率及并发活动对隐私与完整性的实际影响。