{"title":"Adversarially robust sequential hypothesis testing","authors":"Shuchen Cao, Ruizhi Zhang, Shaofeng Zou","doi":"10.1080/07474946.2022.2043050","DOIUrl":null,"url":null,"abstract":"Abstract The problem of sequential hypothesis testing is studied, where samples are taken sequentially, and the goal is to distinguish between the null hypothesis where the samples are generated according to a distribution p and the alternative hypothesis where the samples are generated according to a distribution q. The defender (decision maker) aims to distinguish the two hypotheses using as few samples as possible subject to false alarm constraints. The problem is studied under the adversarial setting, where the data generating distributions under the two hypotheses are manipulated by an adversary, whose goal is to deteriorate the performance of the defender—for example, increasing the probability of error and expected sample sizes—with minimal cost. Specifically, under the null hypothesis, the adversary picks a distribution with cost and under the alternative hypothesis, the adversary picks a distribution with cost This problem is formulated as a non-zero-sum game between the defender and the adversary. A pair of strategies (the adversary’s strategy and the defender’s strategy) is proposed and proved to be a Nash equilibrium pair for the non-zero-sum game between the adversary and the defender asymptotically. The defender’s strategy is a sequential probability ratio test and thus is computationally efficient for practical implementation.","PeriodicalId":48879,"journal":{"name":"Sequential Analysis-Design Methods and Applications","volume":"41 1","pages":"81 - 103"},"PeriodicalIF":0.6000,"publicationDate":"2022-01-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Sequential Analysis-Design Methods and Applications","FirstCategoryId":"100","ListUrlMain":"https://doi.org/10.1080/07474946.2022.2043050","RegionNum":4,"RegionCategory":"数学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q4","JCRName":"STATISTICS & PROBABILITY","Score":null,"Total":0}
引用次数: 0
Abstract
Abstract The problem of sequential hypothesis testing is studied, where samples are taken sequentially, and the goal is to distinguish between the null hypothesis where the samples are generated according to a distribution p and the alternative hypothesis where the samples are generated according to a distribution q. The defender (decision maker) aims to distinguish the two hypotheses using as few samples as possible subject to false alarm constraints. The problem is studied under the adversarial setting, where the data generating distributions under the two hypotheses are manipulated by an adversary, whose goal is to deteriorate the performance of the defender—for example, increasing the probability of error and expected sample sizes—with minimal cost. Specifically, under the null hypothesis, the adversary picks a distribution with cost and under the alternative hypothesis, the adversary picks a distribution with cost This problem is formulated as a non-zero-sum game between the defender and the adversary. A pair of strategies (the adversary’s strategy and the defender’s strategy) is proposed and proved to be a Nash equilibrium pair for the non-zero-sum game between the adversary and the defender asymptotically. The defender’s strategy is a sequential probability ratio test and thus is computationally efficient for practical implementation.
期刊介绍:
The purpose of Sequential Analysis is to contribute to theoretical and applied aspects of sequential methodologies in all areas of statistical science. Published papers highlight the development of new and important sequential approaches.
Interdisciplinary articles that emphasize the methodology of practical value to applied researchers and statistical consultants are highly encouraged. Papers that cover contemporary areas of applications including animal abundance, bioequivalence, communication science, computer simulations, data mining, directional data, disease mapping, environmental sampling, genome, imaging, microarrays, networking, parallel processing, pest management, sonar detection, spatial statistics, tracking, and engineering are deemed especially important. Of particular value are expository review articles that critically synthesize broad-based statistical issues. Papers on case-studies are also considered. All papers are refereed.