/adversarial-attack-on-mtcnn

adversarial patch train by I-FGSM to attack on MTCNN face detection system

Primary LanguagePython

Adversarial-attack-on-mtcnn

Evaluate the AP of four different adversarial-patch pasting methods on two types of faces(no mask , mask on)

Real world attack demo

Video : https://www.youtube.com/watch?v=mJUsIuD5B3Q&feature=youtu.be&ab_channel=%E6%B4%AA%E5%95%86%E7%A8%8B>

Patch pasting

The pasting coordinates are in patch_coordinate.txt file

  • chin
  • forehead
  • frame
  • mouth

Picture

  • The test samples have two types of faces
  • Each type has five different sizes of faces(standing one to five meter away from the camera)

Requirements

  • install mtcnn

    pip install mtcnn

  • change mtcnn's factory.py

    keras.layers -> tensorflow.keras.layers

    keras.models -> tensorflow.keras.models

  • save mtcnn.py's scales into scales.npy

    np.save('scales', scales)