Overview

October 2, 2023 ยท View on GitHub

This repository contains the code to reproduce experiments in the paper:

  • GPU.zip: On the Side-Channel Implications of Hardware-Based Graphical Data Lossless Compression

Preliminaries

  • Required tool: perf, stress-ng
  • Required shared libraries: OpenGL library (-lgl, core profile 3.3), GLFW library (-lglfw).
  • Python virtual environment and package installation:
virtualenv -p python3 venv
source venv/bin/activate
pip install numpy
pip install matplotlib
pip install py-cpuinfo

High level summary

  1. 01-leakage-channel reproduces Table 2, and Figure 2 and 3 in the paper.
  2. 02-memory-stressor reproduces Figure 4 in the paper.
  3. 03-llc reproduces Figure 5 in the paper.
  4. 04-chrome-poc contains the two PoCs in Section 5 of the paper. Reproducing Figure 14 and 15 can be done by uprobing Chrome and then run the PoCs with scripts and driver code similar to the ones in 01-leakage-channel.
  5. poc contains a OpenGL workload that attempts to trigger iGPU graphical data compression.
  6. uprobe-chromium attaches a uprobe to the Math.sqrt() function in Chrome.
  7. i915-tools documents our reverse-engineered proprietary compression algorithms of Intel and AMD iGPUs.

Tested machine

  1. 01-leakage-channel is tested on an Intel i7-8700 (Ubuntu 22.04), i5-1135G7 (Debian 11.6), i7-12700K (Ubuntu 22.04), and AMD Ryzen 7 4800U (Ubuntu 20.04).
  2. 02-memory-stressor and 03-llc are tested on an Intel i7-8700 (Ubuntu 22.04).
  3. 04-chrome-poc is tested on machines in Table 3 of the paper.

All tested machines are bare-metal machines. Running the code on other CPUs will likely require modification to the code (e.g., hardware measurement methodology).

Citation

If you make any use of this code for academic purposes, please cite the paper:

@inproceedings{wang2024gpuzip,
    author = {Yingchen Wang and Riccardo Paccagnella and Zhao Gang and Willy R. Vasquez and David Kohlbrenner and Hovav Shacham and Christopher W. Fletcher},
    title = {{GPU.zip}: On the Side-Channel Implications of Hardware-Based Graphical Data Compression},
    booktitle = {Proceedings of the IEEE Symposium on Security and Privacy (S&P)},
    year = {2024}
}