A Picture is Worth 500 Labels: A Case Study of Demographic Disparities in Local Machine Learning Models for Instagram and TikTok Source Code

This repo contains all resources for the paper, "A Picture is Worth 500 Labels: A Case Study of Demographic Disparities in Local Machine Learning Models for Instagram and TikTok" published at IEEE Symposium on Security and Privacy 2024.

We break up the resources into three categories:

1. Dynamic Analysis Code
2. Static Analysis Resources
3. Datasets

Dynamic Analysis Code

We provide all code written that performs our static analysis. All resources are under the Dynamic_Analysis_Code folder. Within this folder there are three sub-folders: OS_Code, python_scripts, and frida_scripts. Each folder will have a README file explaining their respective contents. The OS-Code folder provides all of the source code changes we made to the operating system as raw C++ files. The full custom ROM is too large to store on GitHub, if you would like the ROM please email me at jwwest@wisc.edu. We are currently working on a custom patching script to deploy those changes, however, we will provide a written tutorial to install the patch yourself for the time being. The python_scripts folder will contain all python related code. This includes the OS logcat parser and the TikTok experiment script. Finally, the frida_scripts folder contains all Frida scripts used. There are three main scripts we provide: our native hooking script, our Instagram experiment script, and any other custom scripts we wrote during our research.

Static Analysis Code

This folder will contain all Apks and native libraries we examined. We will include a tutorial discussing how we decompiled each component.

Datasets

This folder will provide the list of datasets with their links. It will also provide the experimental data we gathered for the experiments. We will also provide all data used that is not public, meaning all sythetic data used.