zskt

Experiments

The code is based on ZSKT.

Run Experiments

General

• Single experiment. Distill from a badnet_grid-backoored teacher with arch WRN-16-2 to student with arch WRN-16-1.

  export CUDA_VISIBLE_DEVICES=0  # specify GPU
  python main.py --dataset=CIFAR10 --teacher_architecture=WRN-16-2 --student_architecture=WRN-16-1 --trigger_pattern=badnet_grid --seeds=3

• Run sweeps. Choose a sweep command below where you can find all hparams. For example,

  wandb sweep sweeps/cifar10_wrn_poi.yml
  # get the `wandb agent <agent code>` from the CLI output.
  wandb agent <agent code>  # this will run one pair of hyper-params from `cifar10_wrn_poi.yml`.

  wandb agent <> can be run in parallel in different processes, which will auto select different params in yml file.

Trigger	Teacher	Student Acc/ASR
	Acc/ASR	ZSKT	ZSKT+ABD	Clean KD
BadNets (grid)	92.1/99.9	71.9/96.9	68.3/0.7	74.6/4.3
Trojan WM	93.8/100	82.7/93.9	78.2/22.5	77.5/11.1
Trojan 3x3	93.4/98.7	80.9/96.8	71.7/33.3	72.9/1.7
Blend	93.9/99.7	77.0/74.4	71.5/23.1	78.0/4.3
Trojan 8x8	93.7/99.6	80.5/57.2	72.6/17.8	75.2/9.3
BadNets (sq)	93.4/97.8	80.8/37.8	77.9/1.9	76.2/9.1
CL	91.2/94.3	76.8/17.5	67.4/10.2	69.4/2.1
Sig	90.5/97.3	77.9/0.0	72.2/0.	77.4/0.
l2_inv	93.9/100	82.0/0.3	70.7/1.9	77.2/1.2
l0_inv	92.4/99.6	72.8/8.3	69.4/0.	79.2/3.7

Distill from poisoned teachers

Evaluate different backdoors with ZSKT.

wandb sweep sweeps/cifar10_wrn_poi.yml
wandb sweep sweeps/gtsrb_wrn_poi.yml

Distill using clean data

# single run
python kd_distill.py --trigger_pattern=badnet_grid --no_log
wandb sweep sweeps/cifar10_wrn_poi_distill.yml
wandb sweep sweeps/gtsrb_wrn_poi_distill.yml

Customization

Add dataset: Edit get_test_loader in zskt/datasets/datasets.py.

Add model: Edit zskt/models/selector.py to add new architecture and pre-trained model paths.

Defense

• CIFAR10

wandb sweep sweeps/cifar10_wrn_poi_defense.yml

• GTSRB

wandb sweep sweeps/gtsrb_wrn_poi_defense.yml