fordevoted/ntu-1091-spml-adversarial_defense

NTU-1091-Security and Privacy of Machine Learning: Adversarial Defense

Overview

the attacker will use adversarial examples (up to epsilon=8 in the L_infinity norm) to attack your model, which is under white-box setting. In this project, I train a robust model for CIFAR-10 that can defend the adversarial examples.

Model

My defense pipeline include image preprocessing, and adversarial training, etc. The detail can be found in the report hw2_r09942066.pdf, please let me know if you need the model weight(email me), then I will send you the download link.

Defense Model

• resnet50

Attack for evaluation

• PGD
• FGSM

Defense technique

• PGD adversarial training
• data pre-processing
• SLQ
• proposed method

Requirement

• pytorch 1.5.0
• python 3
• built-in modules and numpy
• opencv hw2.py has been tested under this env

Usage

For Evaluating

run python hw2.py [image folder]

Testing

cd src/example
python adversarial_testing.py

Ablation Experiment

cd src/example
python ablation_study.py

Generate PGD examples

cd src/example
python test_PGD.py

Result

67.2% accuracy for PGD attack(max epsilon=0.3) under white-box setting

Contact

If there are any problem that make you cannot execute the program, please contact
[email protected] (I have already test under the specify env).