Security in Artificial Intelligence

WINLAB Summer Internship 2023

Advisors: Tianfang Zhang, Changming Li, Hong Li

Group Members: Rut Mehta, Jacob Morin, Ethan Lung, Damon Lin

Project Objective

Artificial intelligence techniques have been widely integrated into mobile and IoT devices, enabling various functionalities based on vision (e.g., face recognition, speech recognition, and speaker identification). The extended pipeline of building deep neural networks (DNN) produces new attack surfaces, such as attacks during the data collection, model training, and model update stages. Recent research studies discovered an effective yet stealthy attack, called a backdoor attack, which trains a hidden trigger pattern into the DNNs. The backdoored DNNs will misclassify an input as an adversary-specified label if the trigger pattern appears, behaving normally otherwise, making it difficult to be detected. Backdoor attacks originate from the image domain, and recent studies have started investigating audio-domain backdoor attacks (e.g., against voice assistant systems). This project aims to study the vulnerabilities of backdoor attacks in the image and audio domains and develop techniques for attack mitigation.

Week 1

• Established project objectives
• Covered and reviewed background information on topics (adversarial attacks, backdoor attacks, etc.)
• ​https://www.orbit-lab.org/attachment/wiki/Other/Summer/2023/SecurityAI/1412.6572.pdf
• Developed a plan of action for the upcoming weeks

Week 2