Skip to content

Latest commit

 

History

History

BLIP_attack

README.md

VLAttack on the BLIP model

[BLIP Paper]

In this repository, we test VLAttack through the VQA task and NLVR task on the VQAv2 and NLVR2 datasets, respectively. We conducted VLAttack on 5K correctly predicted samples. Instructions are shown below:

Pre-trained Model Preparation

Firstly, download the pretrained BLIP model weights (BLIP with ViT-B, 14M) from the BLIP original repository. We use these weights to generate adversarial samples in our work.

Attack VQAv2

  1. Download the VQAv2 dataset from the original website, and then set the vqa_root in ./configs/vqa.yaml

  2. Download the finetuned VQAv2 model weights from the original repo of BLIP. Specifically, the finetuned model weights can be downloaded from here. Don't forget to set the pretrain in ./configs/vqa.yaml with the path of model_vqa.pth.

  3. Find 5K correctly predicted samples using the python prepare_vqa.py command. After running, it will generate right_vqa_list.txt and right_vqa_ans_table.txt, which store the indexes and predictions of correctly predicted samples.

  4. To conduct VLAttack on the VQAv2 dataset, use the python attack_vqa.py command with different --method options shown below:

  • Method Options:

  • Command: Replace METHOD_NAME with your chosen options from above:

    python attack_vqa.py --method METHOD_NAME

Attack NLVR2

  1. Download the NLVR2 dataset from the original website, and then set the image_root in ./configs/nlvr.yaml
  2. Download the finetuned NLVR2 model weights from the original repo of BLIP. Specifically, the finetuned model weights can be downloaded from here. Don't forget to set the pretrain in ./configs/nlvr.yaml with the path of model_base_nlvr.pth.
  3. Find 5K correctly predicted samples using the python prepare_nlvr.py command. After running, it will generate right_nlvr_list.txt and right_nlvr_ans_table.txt, which store the indexes and predictions of correctly predicted samples.
  4. To conduct VLAttack on the NLVR2 dataset, use the python attack_nlvr.py command with above --method options. For example, run below command to conduct VLAttack:
python attack_nlvr.py --method VLAttack