Manipulating Machine Learning - Poisoning Attacks and Countermeasures for Regression Learning
Explore the critical topic of poisoning attacks and countermeasures in machine learning through this 17-minute IEEE conference talk. Delve into the first systematic study of poisoning attacks on linear regression models, examining how attackers can manipulate training data to influence predictive outcomes. Learn about a theoretically-grounded optimization framework designed specifically for linear regression and its effectiveness across various datasets and models. Discover a fast statistical attack requiring limited knowledge of the training process. Gain insights into a new principled defense method offering high resilience against poisoning attacks, complete with formal guarantees and upper bounds on attack effects. Examine the practical implications of these findings through evaluations on realistic datasets from healthcare, loan assessment, and real estate domains.