SaTC: CORE: Small: Collaborative: Learning Dynamic and Robust Defenses Against Co-Adaptive Spammers

在线声誉系统无处不在，客户可以根据人群的评论来评估企业、产品、人员和组织。例如，Yelp和TripAdvisor根据用户评论对餐厅和酒店进行排名，RateMD允许患者评论医生和医院。然而，垃圾邮件发送者可以利用这些系统通过虚假但伪装得很好的评论(垃圾邮件)来误导和操纵缺乏经验的客户。为全面保护客户和诚信企业，部署了先进的垃圾邮件检测技术。尽管如此，智能垃圾邮件发送者仍然可以探测，然后进化以绕过部署的检测器。该项目研究动态和强大的对策，以击败不断演变的垃圾邮件发送者。这项研究将使监管机构能够实施一个更加公平、透明和值得信赖的在线环境，鼓励企业主提供更高质量的产品和服务，而不是虚假的意见，最终让消费者越来越自信地依赖信誉系统来节省金钱、时间甚至生命。该项目将调查针对学习绕过静态检测器的智能垃圾邮件发送者的自适应垃圾邮件检测技术和系统的设计。调查将遵循两个原则：(1)可以通过检测器和垃圾邮件发送者的行为来感知他们的目标和工作；(2)双方都应该动态地采取行动，以最优的方式击败与对方的行为相适应的对手。基于这些原理，研究人员的目标是：(I)调查动态垃圾邮件的足迹，并将所获得的见解形式化到针对静态检测器的规避模型中；(Ii)通过深度强化学习和马尔可夫博弈来建模不断演变的垃圾邮件发送者和动态检测之间的交互；以及(Iii)引入多个协作垃圾邮件发送者，通过多代理和分层强化学习来通知更复杂的垃圾邮件发送者-检测器的共同适应。研究目标将得到衡量标准和评估的补充，这些指标和评估捕捉现实的垃圾邮件发送者和检测器的目标和约束。该项目将为研究社区带来数据集、算法和试验台系统，并将教育软件和材料游戏化，以提高更广泛人群对虚假内容的认识。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Online reputation systems are ubiquitous for customers to evaluate businesses, products, people, and organizations based on reviews from the crowd. For example, Yelp and TripAdvisor rank restaurants and hotels based on user reviews, and RateMDs allows patients to review doctors and hospitals. These systems can however be leveraged by spammers to mislead and manipulate the inexperienced customers with fake but well-disguised reviews (spams). To comprehensively protect customers and honest businesses, advanced spam detection techniques have been deployed. Nonetheless, intelligent spammers can still probe and then evolve to bypass the deployed detectors. This project investigates dynamic and robust countermeasures to defeat the evolving spammers. This research will allow regulatory agencies to enforce a more fair, transparent, and trustworthy online environment, encourage business owners to offer higher quality products and services rather than fake opinions, and ultimately, allow consumers to increasingly rely on the reputation systems confidently to save money, time and even lives. The project will investigate the design of adaptive spam detection technologies and systems against intelligent spammers that learn to bypass static detectors. The investigation will follow two principles: (1) the goals and workings of the detectors and spammers can be sensed through their behaviors; (2) both parties should act dynamically to optimally defeat their opponents who co-adapt with the other's behaviors. Based on these principles, the researchers aim to: (i) investigate the footprint of dynamic spamming and formalize the gained insights into evasion models against static detectors; (ii) model the interactions between the evolving spammer and dynamic detections through deep reinforcement learning and Markov games; and (iii) introduce multiple cooperative spammers to inform more complex spammer-detector co-adaptations through multi-agent and hierarchical reinforcement learning. The research aims will be complemented by metrics and evaluations that capture realistic spammer and detector goals and constraints. The project will result in datasets, algorithms, and testbed system for the research community, and gamified educational software and materials to increase awareness of fake contents among a broader population.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.