Aditya Machiroutu

Project Portfolio

Machine Learning Poisoning with Convex Polytope

Abstract or project description

As the world today diverts further into technology, artificial intelligence and machine learning have become a prominent topic. The heart of all artificial intelligence is its models. AI models serve many purposes, from addressing health concerns and their diagnosis to avoiding dangers like car accidents. Data poisoning, the center of this project, occurs when an AI model is hijacked, leading to faulty data causing incorrect results which ruins the model. For example, certain pixels on an image are changed, which leads a model to misclassify the object. In certain cases, like an AI model unable to detect a stop sign, this misclassification can have serious consequences. In my project, I address these concerns by finding different ways to benchmark attacks. I also find simplified plans of attacking a machine learning model and potential defenses against these attacks.