Predictive Response Optimization - Using Reinforcement Learning to Fight Online Social Network Abuse

Learn how reinforcement learning can revolutionize online social network abuse detection and response through this 12-minute conference presentation from USENIX Security '25. Explore the limitations of traditional binary classification approaches to detecting phishing, spam, fake accounts, and data scraping, and discover why simply labeling content as "benign" or "malicious" falls short of effective abuse mitigation. Examine the Predictive Response Optimization (PRO) system, which moves beyond detection to action selection by utilizing contextual information to predict future abuse and user-experience metrics for each possible response. Understand how this reinforcement learning-based approach optimizes the tradeoff between preventing harm from abuse and minimizing impact on legitimate users by expanding beyond simple ban/allow decisions to include actions like CAPTCHAs and evidence collection. Analyze real-world deployment results from Instagram and Facebook, where PRO demonstrated significant improvements over baseline classification systems, reducing abuse volume by 59% and 4.5% respectively without negatively affecting user experience. Discover how the system adapts automatically to changing business constraints, system behaviors, and evolving adversarial tactics through detailed case studies presented by researchers from Meta Platforms.