Unlearning Attack

This folder contains the VGG retrain attack workflow.

Installation

Install dependencies before running the examples:

pip install -r requirements.txt

Download LlamaFactory from the released repository, then place it under src/LlamaFactory:

git clone <ANONYMOUS_LLAMAFACTORY_URL> src/LlamaFactory

If you receive LlamaFactory as an archive instead, extract it so that the package root is src/LlamaFactory.

One-Command Launch

Run both IND and OOD retrain attacks:

bash run_vgg_retrain.sh

Run one side only:

bash run_vgg_retrain.sh ind
bash run_vgg_retrain.sh ood

Extra arguments are passed to the Python entrypoints:

bash run_vgg_retrain.sh ind --overwrite
METHOD_NAMES="UNDIAL SatImp" bash run_vgg_retrain.sh ood

Included Files

• scripts/vgg/vgg_test_retrain_ind.sh
• scripts/vgg/vgg_test_retrain_ood.sh
• src/test/vgg_test_retrain_ind.py
• src/test/vgg_test_retrain_ood.py
• requirements.txt

External Inputs

By default, the launch scripts look for datasets and models under project-local folders:

• DATA_ROOT=./data
• MODEL_ROOT=./models/unlearn
• INPUT_PARQUET=${DATA_ROOT}/facerec/facerec_test.parquet
• FORGET_PARQUET=${DATA_ROOT}/facerec/qwen2.5_vl_3b_train_forget.parquet
• SFT_PARQUET=${DATA_ROOT}/pacs/qwen2.5_vl_3b_train_1over20.parquet
• MODEL_PATH=${MODEL_ROOT}/rec_Qwen2.5-VL-3B-Instruct_${METHOD_NAME}

Place each already-trained unlearned model under the matching MODEL_PATH directory, for example ./models/unlearn/rec_Qwen2.5-VL-3B-Instruct_mmunlearner.

The PACS SFT parquet above is a 1/20-size subset used as a lightweight example.

You can override them with environment variables before running the script.