Recognition model using Yale Face Datset

• Manual
• Model Architecture
• Algorithm
• Performance & Demo
• Reference
• Evasion Examples

Manual

• Setting
  1. git clone this repo
  2. source bin/activate
  3. pip install -r requirements.txt
• Train a model and Save as model.pt & Test accuracy
  1. python main.py --execTrain 1 --saved 1
     • If don't want to save the model, ignore --saved option
• Demo
  • python demo.py

Model

Original Model

Kernel Size = 5
Stride = 2
Final Output # = 15

Performance & Demo

• Epoch : 50 / Learning Rate : 0.01
• Loss

• Accuracy

• Demo
  • Select random image from original 164 data and predict
  • The normal image of the predicted is shown for compare

Evasion Examples

Model trained with FGM Examples

Create FGM Adversarial Example by using art.attacks.evasion.FastGradientMethod
The training method was carried out as shown in "EXPLAINING AND HARNESSING ADVERSARIAL EXAMPLES".

More Details for CNN Model trained with FGM Examples

Model trained with PGD Examples

Create FGM Adversarial Example by using art.attacks.evasion.ProjectedGradientDescent
The training method was carried out as shown in "Towards Deep Learning Models Resistant to Adversarial Attacks".

More Details for CNN Model trained with PGD Examples

Compare with Original Model

	Original Model	FGM Model	PGD Model
Original Examples	0.882	0.941	0.852
FGM Examples	0.5	0.941	0.941
BIM Examples	0.058	0.941	0.941
PGD Examples	0.029	0.852	0.970

More Details for Performance

Reference

• Ian J. Goodfellow, Jonathon Shlens & Christian Szegedy, "EXPLAINING AND HARNESSING ADVERSARIAL EXAMPLES", ICLR 2015
• Alexey Kurakin, Ian J. Goodfellow, Samy Bengio, "ADVERSARIAL EXAMPLES IN THE PHYSICAL WORLD", ICLR 2017
• Aleksander Ma ̨dry, Aleksandar Makelov, Ludwig Schmidt, "Towards Deep Learning Models Resistant to Adversarial Attacks", stat.ML 4 Sep 2019
• UCSD Computer Vision, Yale Face Database Download
• Adversarial-Robustness-Toolbox Link