CTD-ARTIFACT

This repository contains companion scripts to run artifact evaluation for our Usenix 24 paper.

Hardware dependencies

The CTD model needs proper environment setup to run GPU based inference. Therefore, it is recommended to run the evaluation experiments on environments with following configuration met

1. CUDA Toolkit (v11.3)
2. CUDNN (v8.2)
3. GCC (v9.3.0)

---

GPU memory requirements

The backbone model used by CTD is Google's FLAN-T5-XXL. We use a sharded version of this model loaded in 8-bit configuration to make sure inference is possible in consumer hardware. It is recommended to run the command nvidia-smi on the development environment to make sure appropriate GPU memory requirements are met.

---

Installing Python Packages

After configuring software dependencies, please run pip install -r requirements.txt to install all dependencies are installed. You can run python basic_test.py to make sure the necessary dependencies are met and move forward with the individual experiments

1. experiment1_bootstrapping.py
2. experiment2_evaluating.py
3. experiment3_integrating.py

---

Issues you might run into and workarounds

Issue #1 CUDA OOM error

Workaround: Try decreasing the value of BATCH_SIZE on the Python scripts. For 24G of memory, batch size of 50 should work.

---

Citing The Paper

@inproceedings {sec24:ctd,
    title = {Enabling Contextual Soft Moderation on Social Media through Contrastive Textual Deviation},
    booktitle = {33rd USENIX Security Symposium (USENIX Security)},
    year = {2024},
    author={Pujan Paudel, Mohammad Hammas Saeed, Rebecca Auger, Chris Wells, and Gianluca Stringhini}
}