comeeasy / r320_vonenet

FGSM_MNIST

We're trying to make model better which is robust against to adversarial images, especially made by FGSM. Yann LeCun's MNIST datasets are used.

We're inspired by this tutorial.

Fine-tune modeling

1. train model with original MNIST datasets (learning rate == 0.001)
   
2. get adversarial images of MNIST from trained model
   
3. fine-tune model with adversarial images. learning rate is 0.0001 (it may be modified)
   
4. validate with validation set 100 epochs each models
   
5. results saved as a plot
   

A function named
generate_image_adversarial(args) is just interpretation of tensorflow code to pytorch code

Results

red line : accuracy of original MNIST imagess of fine-tuned model
blue line : accuracy of adversarial MNIST images of fine-tuned model

1. Result of none VOneNet finetuned
   
   
   

2. Result of VOneNet finetuned
   
   
   

fine-tunning harms None-VOneNet models's prediction of original data.

But VOnetNet models are robust to fine-tunning

Requirements

• python 3.8+
• pytorch 0.4.1+
• numpy
• tqdm

License

MIT License

trained model

Name	Description
1-layer-linear-classifier	really simple model
3-layer-linear-classifier	add two layer to 1-layer simple model
Convnet	simple convolutional model

Report

This browser does not support PDFs. Please download the PDF to view it: Download PDF.

Longer Motivation

1. VOneNet maybe boosts performance. So we're considering how apply this model to VOneNet