How to Hack an AI
19SANS AI Cybersecurity Summit 2023
Speaker: Harriet Farlow, CEO, Mileva Security Labs
Adversarial machine learning (or AML) is a field growing in prominence that represents the ability to “hack” Artificial Intelligence (AI) and Machine Learning (ML) algorithms by poisoning data sets imperceptibly before training, by evading classification, leaking confidential information or by hijacking the model's function to make it do something it wasn't intended to. The rapid uptake of AI/ML systems by organizations means the attack surface is growing significantly. I believe AI/ML security may soon join cyber security as one of the greatest technological and geostrategic threats. However, there is still time to learn from the lessons of cyber security.
This talk is intended to inform information security professionals about the increasing relevance of their field to AML and AI/ML security. It will describe how ML models work, why vulnerabilities exist and how they can be exploited. I will demonstrate the cutting edge of AML - glasses that deceive facial recognition detectors, stickers that can disguise objects in the physical world from image classification engines, and how carefully crafted noise can cause speech to text systems to hear messages that humans can't. I will also describe some of my own research. The audience should come away with an appreciation for the field of AML, why AI/ML security is a growing concern, and how in their roles they can contribute to the dialogue.
View upcoming Summits: http://www.sans.org/u/DuS