The code is for paper "Robust Unsupervised Domain Adaptation from A Corrupted Source".
We conduct two kind of unsuerpvised domain ataptation (UDA) tasks: digits tasks (USPS <-> MNIST), and image tasks (CIFAR10 <-> STL)
We utilize two kinds of backdoor attacks to corrupt the source training data: BadNet and CLBD
To download digits datasets, you can e.g. unpack the following dataset in data/digits.
we provide a detailed description on key arguments.
| Arguments | Discription |
|---|---|
| method | UDA method, such as DANN, CDAN |
| task | the argument is only for digits domain adaptation tasks, including minist2usps and usps2mnist |
| corrupt | corruption method for source data |
| poison_ratio | The ratio of poison samples |
| block | the number of blocks for MoM algorithm, if block=1, no MoM algorithm is applied |
| cls_par | this argument is for fine-tune, if cls_par = 0.0, only IM is used, if cls_par > 0, both IM and PL are used |
| s_dset, t_dset | these two arguments are only for image tasks, and s_dset indicates the source domain, while t_dset indicates the target domain |
| poison | binary arguments, 1 for poison, and 0 for clean |
To conduct the MoM method for digits tasks:
python digits_MOM_poison.py CDAN --block 15 --poison 1 --poison_ratio 0.02 --corrupt badnet
To conduct the MoM method for digits tasks with fine-tune:
python digits_MOM_SHOT.py CDAN --block 15 --poison 1 --poison_ratio 0.02 --corrupt badnet --cls_par 0.0
To conduct the MoM method for image tasks:
python cifar_MOM_poison.py CDAN --corrupt clbd --block 10 --poison_ratio 0.5 --num_iterations 20000
To conduct the MoM method for image tasks with fine-tune:
python cifar_MOM_SHOT.py DANN --corrupt clbd --block 10 --cls_par 0.0 --poison_ratio 0.5 --num_iterations 20000