Adversarial Robustness of AI Agents Acting in Probabilistic Environments
Adversarial Robustness of AI Agents Acting in Probabilistic Environments
复制标题
DOI：
--
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Lisa Oakley;Alina Oprea;S. Tripakis
中科院分区：
文献类型：
--
作者：
Lisa Oakley;Alina Oprea;S. Tripakis
随着机器学习系统在安全关键任务中变得越来越普遍，仔细分析其对攻击的鲁棒性非常重要。我们的工作重点是开发一个可扩展的框架，用于验证机器学习系统随着时间的推移的对抗鲁棒性，利用现有的概率模型检查和优化方法。我们目前的初步进展，并考虑未来的方向验证几个关键属性对复杂的，动态的攻击者。
—As machine learning systems become more pervasive in safety-critical tasks, it is important to carefully analyze their robustness against attack. Our work focuses on developing an extensible framework for verifying adversarial robustness in machine learning systems over time, leveraging existing methods from probabilistic model checking and optimization. We present preliminary progress and consider future directions for verifying several key properties against sophisticated, dynamic attackers.