Image: Congratulations to our Data Science student Lauren Rose Wilkes who got selected to the Research Experiences for Undergraduates program at MIT. In Summer 2022, she worked under the supervision of Dr. Aleksander Madry in the Computer Science and Artificial Intelligence Lab. There she investigated machine learning model robustness in image classification by identifying the effect of poisoning the images most influential to the model prediction. Her project leveraged datamodels, a tool that can determine the influence of the training images on each other, to select the most influential points. By poisoning the images that are most influential to other images in the model, this project investigated the possibility of breaking the model using fewer poisoned images. By successfully choosing poisons, the project exposed the ease and risk of data poisoning if someone has access to training data and explored how machine learning models learn patterns. Type of News/Audience: Departmental
