Add the Inception series of models that we use frequently to TransferAttack(Maybe have mistakes just for saving here))
This code is modeled after TransferAttack and tf_to_pytorch_model based on the common model (Inception series) pytorch framework, but the recurrence result is relatively large float, need to be checked and modified, to be improved.
The basic operation is the same as TransferAttack. Some of the differences are explained below:
python main_dev.py --input_dir ./path/to/data --output_dir adv_data/ifgsm/resnet18 --model_dir ./path/to/model_dir/ --attack mifgsm --model=resnet18
python main_dev.py --input_dir ./path/to/data --output_dir adv_data/ifgsm/resnet18 --model_dir ./path/to/model_dir/ --eval
The dataset is changed, which is 1000 299×299 ImageNet images that are commonly used. Torch_nets and Torch_nets_weight can be found at tf_to_pytorch_model.
The result of code reproduction is not ideal, there may be errors. The author will check again if there is time in the future. If you are interested in communicating with me, please contact me: [email protected]， and I would be honored to hear from you. Thanks again to TransferAttack and tf_to_pytorch_model for their contribution. In addition, this code is only for private research, if you want to use it, please contact me for permission.