Stackelberg Games for Adversarial Learning

对抗性机器学习涉及数据挖掘者面临活跃对手攻击的情况。数据挖掘者和对手之间的交互可以建模为两个玩家之间的游戏。虽然一些博弈论模型假设玩家同时行动，但在这个项目中一个更合适的假设是，玩家可以在做出自己的决定之前观察对手的行动。在Stackelberg游戏模型中，玩家的行为是顺序的。该项目的第一个目标是为更好地理解Stackelberg对抗学习问题开发一个理论框架；然后，这一步将为构建快速准确的算法来解决它们的过程提供信息。将考虑各种场景，包括攻击者是领导者而相应的数据挖掘者是追随者的情况，反之亦然。理论框架和由此产生的解决算法将应用于基于开源数据集的实际问题。

Adversarial machine learning concerns the situation where data miners face attacks fromactive adversaries. The interactions between the data miner and the adversary can bemodelled as a game between two players. While some game theoretic models assumethat the players act simultaneously, a perhaps more appropriate assumption in this project is that players can observe their opponents' actions before making their own decision. In the Stackelberg game model, players act sequentially, allowing for such an assumption. The first aim of this project is to develop a theoretical framework for a better understanding of Stackelberg adversarial learning problems; this step will then inform the process of constructing fast and accurate algorithms to solve them. Various scenarios will be considered, including situations where an adversary is a leader while the corresponding data miner is a follower and vice-versa. The theoretical framework and resulting solution algorithms will be applied to practical problems based on open-source data sets.