SMGEA: Serial-Mini-Group-Ensemble-Attack
SMGEA is a new Black-Box Adversarial Attack against various Pixel-to-Pixel Tasks, such as Saliency Detection, Depth Estimation, Image Translation, etc. This code repository is an Open-Source Toolbox based on Pytorch Platform.
A preliminary version of this repository has been accepted by AAAI2020: ‘‘A New Ensemble Adversarial Attack Powered by Long-term Gradient Memories’’
We provide 3 visualizations (GIF format) for your reference.
Each GIF contains two parts:
Part-I: In the begining still frames: the upper-left region is the original clean image, the bottom-left region is the ground-truth output of the clean image, the upper-right region is the guide image, the bottom-right region is the ground-truth output of the guide image
Part-II: In the following dynamic frames: the upper-left region is the crafted adversarial example, the upper-right region is the normalized perturbation (obtained by elemen-wise subtraction of clean image and adversarial example, and normalized by min-max normalization for better obvervation). The bottom regions are the outputs of two black-box target models on the crafted adversarial example. The timestamp denotes the iterations.
Visualizations on LSUN'17 Dataset：
Visualizations on Cityspaces Dataset：
Visualizations on Google Satellite Dataset：
Requirements
-
Pytorch == 3.5.2
-
NVIDIA GPU (at least 16GB memory for ensemble attacks!!)
-
You have to compile the Deformable Convolution Lib by yourself: https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch (Required)
-
If you want to train/design/enhance the victim models from scratch to defend different attacks in our literature, please visit https://github.com/CZHQuality/Sal-CFS-GAN for more details about model training. These two repositories support each other.
Complete version (including feature-space ensemble, long-term gradient auto-update) of our code will be released after our journal version is accepted, thanks!!!