Overview / Usage

Deep Neural Networks have been out in the Wild from quite a long time now. They have achieved human level accuracy in classification, segmentation and the other tasks. Despite of all these things, they can be easily misled or fooled by fabricated images known as Adversarial Images.

Adversarial Images are the kind of images which look similar to the original images but are mis-classified with a very high confidence. For example , the adversarial image of a Panda looks very similar to it but it has some features due to which it is mis-classified by the deep learning model with a very high probability score.

This is a very important field when it comes to production. Let's say we have a deep learning model fitted in an autonomous vehicle whose task is to identify humans. If somehow that model receives an adversarial image of the human, it may classify it as something else, leading to disastrous circumstances.

Study of Adversarial Images can help in making the deep neural networks more robust and adaptive to different conditions thereby strengthening their position in production which is of utmost priority for any industrial problem.

Methodology / Approach

For generating the Adversarial Images, I have adapted an InfoGAN like architecture where I exploit the latent features of the Images to generate their adversarial twin.

Entire Pipeline:

1. I first take an Image from the dataset and pass it through the existing models like VGG etc for feature extraction.

2. Then I add some noise to the obtained features and pass it through InfoGAN to generate the adversarial Image.

3. I use MSE( Mean Squared Error) and adversarial Loss( Real/Fake Loss) to train the GAN