Gradient-Aligned Attack

Description

This is the code repository of the paper "Improving Query Efficiency of Black-Box Attacks via the Preference of Deep Learning Models".

Datasets

• ImageNet
• Real-world API, IMAGGA

Pretrained Models

The used surrogate and victim models are the pretrained models from pytorch.org and huggingface.co.

Running Scripts

# Running GAA on the ImageNet in the low-query scenario
python -u ./ImageNet/run_GAA.py --ensemble-surrogate1 --victim-model-type vgg19 --test-batch-size 1 --max-query 5 --gpu 0

# Running BASES with GACE on the ImageNet
python -u ./ImageNet/run_BASES_with_GACE.py --ensemble-surrogate1 --victim-model-type vgg19 --test-batch-size 1 --gpu 0

# Attacking real-world API, IMAGGA
python -u ./ImageNet/attack_IMAGGA.py --ensemble-surrogate1 --test-batch-size 1 --gpu 0

Note that

The predict_status_arr_v2.txt file in the ImageNet/saved_model folder needs to be replaced; it stores information about whether each sample in the used dataset has been correctly classified by all the surrogate and victim models.