Attack on Image Classification Network

This project works in two-stage:-

1. Classification of Image from ImageNet dataset using VGG-16 and Inception-V3.

2. Performing attack on the image (adding some perturbation to the original image so that the same model classify the image incorrectly)

Prerequisites

Python, TensorFlow, Foolbox, Download ImageNet dataset. Use Google Colab for running the project if you don't have high computing power.

Source:-

https://arxiv.org/pdf/1801.00553.pdf

Classification Model

Pre-trained model VGG16 and Inception-v3 for classification of an image from IMAGENET dataset.

Attack

1. LBFGS Attack:- This attack has been performed directly by importing LBFGS attack from foolbox library.

2. Iterative FGSM Attack:- This attack has been performed by computed gradient iteratively to compute adversarial input.

3. PGD(Projected Gradient Descent) Attack:- This attack has been performed by iterative computation of gradient descent in such a way that the difference between given input and adversarial input always remains within a valid range