Machine Learning Models Gone Rogue - How to Reel Them Back In

Monica Kay Royal opens by arguing that machine learning models are trained on sensitive personal data and can themselves become a privacy risk. She draws on the book Weapons of Math Destruction and a Cornell paper showing that models can memorize more of their training data than intended. She then walks through credit underwriting, targeted advertising and hiring, describing the personal signals each one draws on.

The middle of the talk covers defenses. She explains masking, generalization and pseudonymization as ways to remove identifiers, then describes how the Cornell researchers reconstructed training images from models using white-box and black-box attacks. Countermeasures include data minimization, adding noise through perturbation, and synthetic data. She also introduces k-anonymity, l-diversity and t-closeness for group privacy, noting the tradeoff with data utility.

The closing section turns to governance. Following the book's argument, she suggests revisiting success metrics that reward only profit and efficiency, adding feedback loops and transparency, and running algorithmic audits. She flags regulatory gaps, such as personality tests used in hiring, and calls for cross-disciplinary review involving law, ethics and audit. The talk ends with resources and a plug for her LinkedIn Learning course.