Lean, Data-Driven Social Media Security

Philip Tully

Social media has ushered in an era of rapid and widespread access to information, but its afforded conveniences come not without potential risks. While digital communication is as easy as ever, the information being communicated can just as easily comprise abusive and even malicious content. From malware, to phishing links, to financial scams, to fake news, to spam botnets - the landscape of social threats facing users is just as diverse and continuously evolving as the networks themselves. Although human experts can distinguish threatening from benign content, the scale of social data demands more statistical methods that are robust to adversarial drift.

To address these concerns, I’ll introduce a flexible machine learning workflow for classifying social network-agnostic text, image and behavioral data. Using real-world examples, I’ll detail how attacker patterns can be learned in order to predict new and incoming threats. Availability of social data is also useful for red team simulations, and I’ll explain how traditionally manual attack workflows like spear phishing and steganography can be automated using machine learning. Through the lens of these different approaches, I’ll show how security data practitioners can remain agile by aligning the batch-driven software development life cycle with the interrupt-driven nature of threat research.