This is the official repository of our NeurIPS 2024 paper: Intruding with Words: Towards Understanding Graph Injection Attacks at the Text Level.

Setup

Run the following command to install dependencies.

pip install -r requirements.txt

Create necessary directories:

mkdir atkg/
mkdir openai_io/
mkdir openai_io/db

Change the api key in api.py to your own.

def efficient_openai_text_api(input_text, filename, savepath, sp, ss, api_key="your-key", rewrite = True)

ITGIA

Embedding-level Attacks

cd ./ITGIA_scripts
mkdir -p ../atkg/gtr_norm    # The directory to save GTR-embedding attacked graphs
bash run_cora.sh          # Default dir: atkg/gtr_norm

To replicate the results combining with HAO, change the weights in attacks/attack.py, And run the Embedding-level Attacks. For example, to set HAO weight=3, you should set:

pred_loss += homo_loss * 10 * 3 # In gia_update_features()
pred_loss += homo_loss * 3 # In smooth_update_features()
cd ./ITGIA_scripts
mkdir ../atkg/gtr_norm30    # The directory to save GTR-embedding attacked graphs
# modify the default dir in run_cora.sh to atkg/gtr_norm30
bash run_cora.sh          # Current dir: atkg/gtr_norm30

Embedding to Text

# ITGIA
bash eval_cora.sh

WTGIA

Change the model_path and api_key in eval_cora.sh to your own.

model_path='meta-llama/Meta-Llama-3-8B' # or your local model path here
api_key='sk-' # your api_key here

# Embedding-level Attacks

cd ./WTGIA_scripts
mkdir -p ../atkg/bow         # The directory to save BoW-embedding attacked graphs
bash run_cora.sh          # Default dir: atkg/bow

# WTGIA

bash eval_cora.sh llama     # args: gpt/gpt_topic/llama/llama_topic/llama_mask/llama_topic_mask

VTGIA

# Embedding-level Attacks
cd ./VTGIA_scripts
bash run_cora_mixing.sh   # VTGIA with mixing prompt

Acknowlegement

The repo is based on GIA-HAO and Graph-LLM.