arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2610.15847cs.ITmath.IT
收藏

选择性隐私下的残差信息泛化界:扩展版本

Residual-Information Generalization Bounds Under Selective Privacy: Extended Version

KTH Royal Institute of Technology(瑞典皇家理工学院);UNSW Canberra(新南威尔士大学堪培拉校区)

Amirreza Zamani, Parastoo Sadeghi, Mikael Skoglund

AI总结:

针对选择性隐私下残差信息泛化界问题,通过链式规则转换不匹配约束得到闭式通用证书,结合互信息泛化推导受残差任务不确定性和归一化隐私泄漏支配的界,还得到互补改进、识别障碍及泄漏允许量,证明特定族的可达性并给出自适应应用。

摘要:

信息论泛化界通常控制对任务样本的依赖,而选择性隐私可能仅约束不同的私有坐标。我们研究独立同分布对$(X,Y)$,其中学习者观测$(X^n,Y^n)$,泄露关于$X^n$的信息最多为$\epsilon_n$纳特,并在$Y$上产生损失。链式规则转换将这种不匹配约束转化为闭式通用证书$I(W;Y^n)\leq\min\{nH(Y),nH(Y\mid X)+\epsilon_n\}$。将该证书与互信息泛化结合,得到由残差任务不确定性和归一化隐私泄漏支配的界。我们推导互补的任务压缩和单个样本改进,识别出一个尖锐障碍,表明为何仅选择性隐私不一定能得到消失的证书,并对该界求逆以获得规定泛化容差下的精确泄漏允许量。我们证明确定性任务和独立残差源族的精确信息水平可达性,并表明所得$\sqrt{\epsilon_n/n}$泛化率在通用常数范围内是紧的。我们还给出了固定时域自适应应用,其中条件泄漏成本会累加。

英文摘要:

Information-theoretic generalization bounds usually control dependence on the task sample, whereas selective privacy may constrain only a different, private coordinate. We study i.i.d. pairs $(X,Y)$, where a learner observes $(X^n,Y^n)$, leaks at most $ε_n$ nats about $X^n$, and incurs loss on $Y$. A chain-rule transfer converts this mismatched constraint into the closed-form universal certificate $I(W;Y^n)\leq\min\{nH(Y),nH(Y\mid X)+ε_n\}$. Combining the certificate with mutual-information generalization yields a bound governed by residual task uncertainty and normalized privacy leakage. We derive complementary task-compression and individual-sample refinements, identify a sharp obstruction showing why selective privacy alone need not give a vanishing certificate, and invert the bound to obtain the exact leakage allowance for a prescribed generalization tolerance. We prove exact information-level achievability for deterministic-task and independent-residual source families and show that the resulting $\sqrt{ε_n/n}$ generalization rate is tight up to a universal constant. We also give a fixed-horizon adaptive application in which conditional leakage costs add.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑