Skip to content

Latest commit

 

History

History
112 lines (88 loc) · 5.12 KB

README.md

File metadata and controls

112 lines (88 loc) · 5.12 KB

False-Claims-Against-Model-Ownership-Resolution

Pytorch implement of the paper False Claims against Model Ownership Resolution which is to appear in USENIX Sec '24. https://arxiv.org/pdf/2304.06607.pdf

Dependencies

For Adi and EWE, we use the Watermark-Robustness-ToolBox to reproduce their results; for Li (b), DAWN , and DI, we use their open-sourced implementations; as the source code of Lukas is based on TensorFlow (whereas all others are on Pytorch). We re-implemented their schemes in Pytorch.

We evaluate our methods on Ubuntu 20.04.2 LT with one NVIDIA-V100 GPU. One GPU is required for this implementation.

To install dependencies

pip install -r requirements.txt

And we use the mlconfig to pass configuration parameters to each script. The configuration files used in our paper can be found in configs/.

The models for CIFAR10 and ImageNet can be downloaded from this link. Save the corresponding models in checkpoint/ like checkpoint/cifar10. The ten classes dataset of ImageNet we use are reported in our research report. It can also be downloaded from the link above. Save the small_imagenet file to ~/data/. The downloaded files structure should be the follwoings:

  • Falseclaims/
    • checkpoint/
      • cifar10/
        • cifar_wide_resnet/
        • ind/
        • victim/
      • imagenet/
        • ind/
        • resnet18_imagenet/
        • victim/
    • defences/
      • cifar10/
      • imagenet/

Claim Generation and Verification

If the files are downloaded, you can run the commands in commands.txt to generate the false claims. The results are listed in Table~4 in our paper, which is an average of the MORacc reported for the five models we use in experiments. For columns Adi, Li(b), and Lukas, the MORaccs reported are the untargeted results, for columns EWE and DAWN, the MORaccs reported are the targeted results. For the column DI, run the jupyter notebook in defence/.

To train your own models which are to generate the adversarial examples

python train.py -c ['dir to the configure file']

For CIFAR-10

python train.py -c configs/cifar10/train/resnet.yaml

For ImageNet

python train.py -c configs/imagenet/train/resnet.yaml

To generate our false claims and test the MOR accuracy:

python attack.py --c ['dir to the configure file']

For CIFAR-10

python attack.py --c configs/cifar10/train/resnet.yaml

For ImageNet

python attack.py --c configs/imagenet/train/resnet.yaml

To generat false claims with different structures and the same data

For CIFAR-10

python attack.py --c configs/cifar10/train/resnet.yaml --adv False

For ImageNet

python attack.py --c configs/imagenet/train/resnet.yaml --adv False

To generate false claims with the same structure and different data

For CIFAR-10

python attack.py --c configs/cifar10/train/resnet_same_struct.yaml

For ImageNet

python attack.py --c configs/imagenet/train/resnet_same_struct.yaml

To generate the targeted false claims, add `adv_target True', for example, to generate targeted false claims with the same structure and different data for CIFAR-10

python attack.py --c configs/cifar10/train/resnet_same_struct.yaml --adv_target True

To train the independent models and victim models

python attack.py --c ['configs'] --ind_resume False --inds_resume False --vic_resume False

This command will generate the independent models and vicim mdoels then generate our false claims.

To compute the decision thresholds, for Li (b), DAWN, Lukas and DI, we recommend their official implements; and Watermark-Robustness-ToolBox for Adi and EWE.

We also have a simple implementation respectively for Li(b), DAWN, Lukas and DI in the defence/ to compute the decision thresholds.

Note that for Lukas, we first train surrogate models.

python train_surrogate.py -c configs/cifar10/train/resnet_lukas.yaml

The false claims results

Durining the experiments, there will be some logs saving in the directory. For model training logs, they will be saved in /log'. And the images of adversarial examples we generated will be saved in /exp_log`, it shows the influence of eps on images visually.