Aman Priyanshu, Yash Maurya, and Vy Tran, Carnegie Mellon University
Despite the growing reliance on differential privacy to shield user data in interest-based advertising, critical gaps remain in our understanding of its effectiveness against sophisticated threats. This presentation zeroes in on Privacy Attacks, highlighting their significance in truly appraising privacy in such settings, and explores whether LLMs/LMs could serve as formidable attackers. We will explore Google's Topics API, a pioneering effort to balance user privacy with advertising needs, to identify and quantify its vulnerabilities to re-identification and membership inference attacks. Leveraging practical simulations, we expose how edge cases and niche topics within the API amplify re-identification risks, a concern underexplored in prior literature. Our use of Large Language Models (LLMs) to simulate attacks marks a significant departure from traditional analysis, uncovering a heightened accuracy in user re-identification that challenges the API's privacy assertions. The findings underscore a pressing need for the PETs community to pivot towards evaluating the resilience of privacy technologies against LLM-driven threats, ensuring that mechanisms like the Topics API can truly withstand the evolving landscape of digital privacy risks.
Authors:
Aman Priyanshu (Carnegie Mellon University) apriyans@andrew.cmu.edu
Yash Maurya (Carnegie Mellon University) ymaurya@andrew.cmu.edu
Suriya Ganesh Ayyamperumal (Carnegie Mellon University) sayyampe@andrew.cmu.edu
Vy Tran (Carnegie Mellon University) vtran@andrew.cmu.edu
Saranya Vijayakumar (Carnegie Mellon University) saranyav@andrew.cmu.edu
Hana Habib (Carnegie Mellon University) htq@andrew.cmu.edu
Norman Sadeh (Carnegie Mellon University) sadeh@cs.cmu.edu
Aman Priyanshu, Carnegie Mellon University
Aman Priyanshu is a master's student at Carnegie Mellon University, specializing in Privacy Engineering. He is currently working under Professor Norman Sadeh and Professor Ashique KhudaBukush on AI for Social Good. Aman has earned recognition as an AAAI Undergraduate Scholar for his work in Fairness and Privacy. His professional experience includes working at Eder Labs R&D Private Limited, Concordia University (MITACS Globalink Research Scholar), and the Manipal Institute of Technology.
Yash Maurya, Carnegie Mellon University
Yash Maurya is a Privacy Engineering graduate student at Carnegie Mellon University, aiming to develop AI solutions that prioritize privacy and ethics. His work focuses on creating systems that safeguard societal values, with interests in Federated Learning, Differential Privacy, and Explainable AI. He is currently working with Professor Virginia Smith, on Unlearning in LLMs. Yash is also working as a Research Assistant, building a user-centric notice and choice threat modeling framework.
Vy Tran, Carnegie Mellon University
Vy Tran, currently a second-year undergraduate at Carnegie Mellon University, majors in Information Systems and minors in Information Security, Privacy and Policy. Vy is passionate about integrating privacy and security into both physical and digital domains and exploring the ML-privacy nexus alongside her graduate peers. She is set to intern in summer 2024 with The Washington Post's Cyber Security & Infrastructure team, and in summer 2025 with PwC's Cyber Defense & Engineering Consulting team.
author = {Aman Priyanshu and Yash Maurya and Vy Tran},
title = {Through the Lens of {LLMs}: Unveiling Differential Privacy Challenges},
year = {2024},
address = {Santa Clara, CA},
publisher = {USENIX Association},
month = jun
}