SI-AI-FGSM

🚧 WIP

Boosting adversarial attack with AdaGrad, AdaDelta, RMSProp, Adam and more...

Requirements

• Python 3.6.5
• Tensorflow 1.12.0
• Numpy 1.15.4
• opencv-python 3.4.2.16
• scipy 1.1.0

Experiments

Download the data and pretrained models

Running

• python agi_fgsm.py to generate adversarial examples for inception_v3 using AGI-FGSM;
• python ri_fgsm.py to generate adversarial examples for inception_v3 using RI-FGSM;
• python ai_fgsm.py to generate adversarial examples for inception_v3 using AI-FGSM;
• python si_agi_fgsm.py to generate adversarial examples for inception_v3 using SI-AGI-FGSM;
• python si_ri_fgsm.py to generate adversarial examples for inception_v3 using SI-RI-FGSM;
• python si_ai_fgsm.py to generate adversarial examples for inception_v3 using SI-AI-FGSM;
• python si_agi_ti_dim.py to generate adversarial examples for inception_v3 using SI-AGI-TI-DIM;
• python si_ri_ti_dim.py to generate adversarial examples for inception_v3 using SI-RI-TI-DIM;
• python si_ai_ti_dim.py to generate adversarial examples for inception_v3 using SI-AI-TI-DIM;
• python simple_eval.py: evaluate the attack success rate under 8 models including normal training models and adversarial training models.

Acknowledgements

Code refers to NAG attack