Awesome

robust-model-watermarking

This is the official implementation of our paper Towards Robust Model Watermark via Reducing Parametric Vulnerability, accepted by the International Conference on Computer Vision (ICCV), 2023. We will release our codes soon.

Train vanilla watermarked models

To watermark the model using "Content" watermark samples, run the following code. CUDA_VISIBLE_DEVICES=0 python train.py --seeds=[0,1,2] --method=STD --model=ResNet18 -wt=test

Checkpoints and and training logs can be found in dfs/cifar10_y0_test_40000_400_c1.00e+00/STD_wd5.00e-04/ResNet18

To use "Noise" or "Unrelated" watermark samples, replace -wt=test with -wt=gauss -t=0.1 or -wt=svhn respectively

Train APP watermarked model

CUDA_VISIBLE_DEVICES=0 python train.py --seeds=[0,1,2] --method=APP --alpha=1e-2 --app-eps=2e-2 --model=ResNetCBN18 -wt=test

Checkpoints and and training logs can be found in dfs/cifar10_y0_test_40000_400_c1.00e+00/APP_a1.00e-02_rl2_eps2.00e-02_pbs64_bbs64_wd5.00e-04/ResNetCBN18

Evaluate the robustness of the watermarked models

CUDA_VISIBLE_DEVICES=0 python attack.py --seeds=[0,1,2] --method=FT --name=FT_E30_LR5E-2_DROP --ft-lr=5e-2 --ft-lr-gamma=0.5 --ft-lr-drop=[5,10,15,20,25] --ft-max-epoch=30 --target-dir=RESULT_DIR

CUDA_VISIBLE_DEVICES=0 python attack.py --seeds=[0,1,2] --method=FP --name=FP_FINE_E30_LR5E-2_DROP --ft-lr=5e-2 --ft-lr-gamma=0.5 --ft-lr-drop=[5,10,15,20,25] --prune-rate='np.arange(0.8,1,0.05)' --ft-max-epoch=30 --target-dir=RESULT_DIR

CUDA_VISIBLE_DEVICES=0 python attack.py --seeds=[0,1,2] --method=ANP --name=ANP_E30 --anp-max-epoch=30 --target-dir=RESULT_DIR

CUDA_VISIBLE_DEVICES=0 python attack.py --seeds=[0,1,2] --method=NAD --name=NAD_TS2E-2_E10 --ft-max-epoch=10 --ft-lr=2e-2 --ft-batch-size=64 --ft-lr-drop=[2,4,6,8] --nad-max-epoch=10 --nad-lr=2e-2 --nad-batch-size=64 --nad-lr-drop=[2,4,6,8] --target-dir=RESULT_DIR

CUDA_VISIBLE_DEVICES=0 python attack.py --seeds=[0,1,2] --method=MCR --name=MCR_E100_LR1E-2_E50_LR5E-2DROP --ft-max-epoch=50 --mcr-lr=1e-2 --mcr-max-epoch=100 --ft-lr=5e-2 --ft-lr-drop=[10,20,30,40] --target-dir=RESULT_DIR

CUDA_VISIBLE_DEVICES=0 python attack.py --seeds=[0,1,2] --method=NNL --name=NNL --nc-epoch=15 --ft-max-epoch=15 --ft-lr=2e-2 --ft-lr-drop=[10,] --target-dir=RESULT_DIR

Replace the "RESULT_DIR" with the directory of the checkpoints, like dfs/cifar10_y0_test_40000_400_c1.00e+00/STD_wd5.00e-04/ResNet18

The attack results can be found in "RESULT_DIR"