ViT_Inversion

Settings

$ pip install -r requirements.txt

Load the pretrained moco models for GradVit (gv)

• download moco_v2_200ep_pretrain.pth.tar file on the link [(https://huggingface.co/sotaBrewer824/x4ssl/tree/main)] and move it on the demodata folder.

Run Inversion Attacks on the ViT models.

$ python run_attack_imagenet.py --arch vit -a gv --lr 0.1 --max-iters 20000 -b 8 -g 1 
$ python run_attack_imagenet.py --arch vit -a gi --lr 0.1 --max-iters 20000 -b 8 -g 1 
$ python run_attack_imagenet.py --arch vit -a idlg --lr 0.1 --max-iters 20000 -b 8 -g 1 
$ python run_attack_imagenet.py --arch vit -a dlg --lr 1.0 --max-iters 20000 -b 8 -g 1
$ python run_attack_imagenet.py --arch vit -a gs --lr 1.0 --max-iters 20000 -b 8 -g 1
$ python run_attack_imagenet.py --arch vit -a cpl --lr 1.0 --max-iters 20000 -b 8 -g 1

A list of attacks we can do

attack mode 'dlg', 'idlg', 'gs', 'cpl', 'gi', 'gv'

• 'Deep Leakage from Gradients' [DLG]
• iDLG: Improved Deep Leakage from Gradients [iDLG]
• See through Gradients: Image Batch Recovery via GradInversion [gi]
• Inverting Gradients - How easy is it to break privacy in federated learning? [gs]
• GradViT: Gradient Inversion of Vision Transformers [gv]

A list of model architecture we can do

• ResNet18
• ResNet50
• LeNet
• ConvNet
• Vision Transformer