Tomer Laor, Yossi Oren

Poster: LockedApart: Faster GPU Fingerprinting Through the Compute API

Introduction

Poster: lockedapart: faster gpu fingerprinting through the compute api. LockedApart leverages WebGPU for significantly faster (310x) and more accurate (1.8x) GPU fingerprinting. Explore the crucial security implications of the WebGPU Compute API.

Abstract

WebGL offers website direct access to the GPU, allowing beautiful graphics. The direct hardware access offered by WebGL was also shown to expose multiple security vulnerabilities. In particular, DrawnApart showed that by performing graphical micro-benchmarks on the GPU, it is possible to fingerprint the underlying hardware. Recently, the access of websites to the GPU was extended with the introduction of the WebGPU API, a new low-level API that allows websites to perform general-purpose computations on the GPU. Our research question was: Does this additional access to the GPU expose additional avenues for fingerprinting? Our initial results show that this is true. Our new attack, which we call LockedApart, uses WebGPU to directly measure contention between GPU threads. Compared to DrawnApart, LockedApart is up to 310x faster and up to 1.8x more accurate. These preliminary results show that it is important to consider the security implications of this new API. The code for LockedApart is available at https://github.com/LockedApart/LockedApart

Review

This submission presents a timely and relevant investigation into the security implications of the newly introduced WebGPU API, specifically concerning hardware fingerprinting. Building upon previous work like DrawnApart, the authors introduce "LockedApart," a novel attack vector that leverages WebGPU's low-level access to the GPU compute API. The core research question—whether WebGPU exposes new avenues for fingerprinting—is directly addressed with clear initial findings. The paper highlights a critical privacy concern associated with modern web technologies that grant increasing access to underlying hardware, positioning this work as an important contribution to the web security landscape. The methodology of LockedApart is particularly compelling, utilizing WebGPU to directly measure contention between GPU threads, a technique that significantly advances the state-of-the-art in GPU fingerprinting. The quantitative results are impressive: LockedApart is demonstrated to be up to 310 times faster and 1.8 times more accurate than its predecessor, DrawnApart. These improvements are not merely incremental; they indicate a substantial leap in the practicality and potency of GPU-based fingerprinting attacks. The provision of the attack code on GitHub further enhances the work's credibility and facilitates independent verification and future research. The findings have important implications for web privacy and the secure design of future web APIs. While the authors state these are "preliminary results," they unequivocally underscore the need for developers and browser vendors to consider the security and privacy ramifications of WebGPU. Future work could delve into potential mitigation strategies against such attacks, explore the generalizability of LockedApart across a wider range of GPU architectures, and analyze the feasibility of deploying such an attack in real-world scenarios. This paper represents a significant initial step in understanding and addressing a looming threat, and its contribution merits attention from the research community.

Full Text

Comments

You need to be logged in to post a comment.