Complementary Label Model Attack(CLMA)

Preface

Since our reverse complementary label model (RCLM) is trained on PyTorch, the parameters of RCLM are saved as .pt file, so our attack experiment is divided into two parts.

• Fist step: With the PyTorch environment, save the gradient of internal feature as weights of the CLMA section.
• Second step: With the TensorFlow environment, load the weights of CLMA, combine with the weights of FIA, execute the subsequent attack progress.

Requirements

TensorFlow Environment

• Python 3.8.0
• Keras 2.4.3
• Tensorflow 2.5.0
• Numpy 1.21.2
• Pillow 8.4.0
• Scipy 1.2.1
• Torch 1.9.0

PyTorch Environment

• Python 3.7.0
• Torch 1.10.2
• Torchvision 0.11.3
• Numpy 1.21.2
• Pillow 6.2.1

Experiments

Introduction

• attack.py : the implementation for different attacks.

• verify.py : the code for evaluating generated adversarial examples on different models.

  You should download the pretrained models from ( https://github.com/tensorflow/models/tree/master/research/slim, https://github.com/tensorflow/models/tree/archive/research/adv_imagenet_models) before running the code. Then place these model checkpoint files in ./FIA/models_tf.

Example Usage

Generate and save CLMA weights:

• CLMA

python CLMA.py

Some parameters need to be adjusted inside CLMA.py, as explained in the annotation.

Generate adversarial examples:

• FIA

python attack.py --model_name vgg_16 --attack_method FIA --layer_name vgg_16/conv4/conv4_3/Relu --ens 30 --probb 0.7 --output_dir ./adv/FIA/ --beta 0

• PIM

python attack.py --model_name vgg_16 --attack_method PIM --amplification_factor 10 --gamma 1 --Pkern_size 3 --output_dir ./adv/PIM/

• FIA+CLMA

python attack.py --model_name vgg_16 --attack_method FIA --layer_name vgg_16/conv4/conv4_3/Relu --ens 30 --probb 0.7 --output_dir ./adv/FIA/ --beta 0.05

In the TensorFlow environment, our approach CLMA can be combined with FIA only by changing the beta (>0).

Evaluate the attack success rate

python verify.py --ori_path ./dataset/images/ --adv_path ./adv/FIA/ --output_file ./log.csv