GZHU-DVL / DMA

Code for paper "Boosting Query Effificiency of Meta Attack with Dynamic Fine-tuning"

Dynamic Meta Attack

---

Boosting Query Effificiency of Meta Attack with Dynamic Fine-tuning

Da Lin, Yuan-Gen Wang, Senior Member, IEEE, Weixuan Tang, Member, IEEE,

Xiangui Kang, Senior Member, IEEE

IEEE Signal Processing Letters, 2022

Setup

Requirements

• Pytorch (torch = 1.7.1, torchvision = 0.8.2) packages
• Python 3.6

We evaluate the proposed method on the MNIST, CIFAR10 datasets.

CIFAR-10 Experiment

untargeted attack

cd DMA_cifar
python test_all.py

targeted attack

cd DMA_cifar
python test_all.py --untargeted False

attack with randomly initialized meta attacker

Please delete meta_model.load_state_dict(pretrained_dict) in line 58 of load_attacked_and_meta_model.py first.

cd DMA_cifar
python test_all.py

License

This source code is made available for research purposes only.

Acknowledgment

Our code is built upon MetaAttack_ICLR2020.