model_extraction_interiit

Training and testing the attacker model with KL divergence loss

Use this command to train and test the attacker model in the vanilla strategy of backpropagating KL divergence loss through attacker model

python run_attacker.py \
  --train_input_dir data/ \
  --train_logits_file logits.pkl \
  --test_input_dir test_data/ \
  --test_logits_file test_logits.pkl \
  --attacker_model_name resnet-lstm \
  --victim_model_name swin-t \
  --epochs 20 \
  --resnet_lstm_trainable_layers 1 \
  --learning_rate 1e-3 \
  --load_test_from_disk

where
train_input_dir is the training data directory,
test_input_dir is the testing data directory,
train_logits_file are the logits generated by victim model on training data
test_logits_file are the logits generated by victim model on testing data (Kinetics)
attacker_model_name is swin-t or movinet
victim_model_name is resnet-lstm or c3d, etc.
resnet_lstm_trainable_layers is specific to resnet-lstm and is the number of unfrozen layers of resnet used
load_test_from_disk is to not load more than a batch of images in memory at a time in testing
load_train_from_disk is to not load more than a batch of images in memory at a time in training