DP-SGD
PulseAugur coverage of DP-SGD — every cluster mentioning DP-SGD across labs, papers, and developer communities, ranked by signal.
4 天有情绪数据
-
新框架为审计RDP机器学习提供最优保证
研究人员开发了一个新的审计框架，用于审计声称具有Rényi差分隐私（RDP）的机器学习算法。该框架使用Donsker-Varadhan（DV）估计器直接测量Rényi散度，为RDP审计提供明确的置信区间。所提出的方法实现了信息论最优的样本复杂度保证，并在经验上优于现有的黑盒方法，尤其是在具有挑战性的小和中等Rényi阶数方面。
-
New theory bounds KAN training, reveals privacy-utility gap
Researchers have established new theoretical bounds for training Kolmogorov-Arnold Networks (KANs), a structured alternative to standard MLPs. The work analyzes KANs trained with mini-batch stochastic gradient descent (…
-
New DP-LAC method enhances private federated LLM fine-tuning
Researchers have developed DP-LAC, a new method for differentially private federated fine-tuning of language models. This technique improves upon existing adaptive clipping methods by estimating an initial clipping thre…
-
New DP-SGD subsampling methods offer improved privacy-utility trade-offs
Two new research papers explore optimized subsampling techniques for Differentially Private Stochastic Gradient Descent (DP-SGD). The first paper, focusing on random shuffling, provides tight upper and lower bounds with…
-
研究人员揭示供应链攻击可窃取本地 LLM 微调中的秘密
研究人员开发了一种新颖的方法，通过利用本地微调的大型语言模型（LLM）供应链代码中的漏洞来窃取敏感信息。该技术超越了被动权重投毒，实现了主动执行劫持，使模型能够记住并泄露特定的秘密，如 API 密钥或个人标识符。该攻击在窃取秘密方面实现了超过 98% 的准确率，同时不影响模型的首要功能，并能绕过 DP-SGD 和代码审计等常见防御措施。