Adversarial Lab

Source code for the website and the project to generate adversarial examples to fool common Machine Learning models.

This is the repository for Visualizing and Comparision of Various Adversarial Attacks on user Uploaded Images using a User-Friendly interface using the DNN framework Pytorch, using popular SOTA Pretrained TorchVision ModelZoo. The Following Attacks have been implemented so far and code can be found inside GAE/attacks.py

1. FGSM
   • Fast Gradient Sign Method, Untargeted
   • Fast Gradient Sign Method, Targeted
2. Iterative
   • Basic Iterative Method, Untargeted
   • Least Likely Class Iterative Method
3. DeepFool, untargeted
4. LBFGS, targeted

Coming Soon: Carlini-Wagner l2, and Many More

Installation

Clone the git repository :

git clone https://github.com/dsgiitr/adversarial_lab.git

Python 3 with Pytorch 1.4.0 is the primary requirement. The requirements.txt file contains a listing of required Python packages; to install all requirements, run the following:

pip3 install -r requirements.txt

Deploying webserver:

After downloading the repo, run flask run:

$ cd adversarial_lab
$ flask run

Fire up your browser and navigate to localhost:5000 or your_server_url:5000. Upload any image in JPG format, Select SOTA torchvision model and Adversarial Attack strategy. Experiment with the parameters for a particular algorithm and push 'Generate'. After a short while, the server returns Ajax response with Perturbed Image and Perturbation of the Original Image along with Top 5 Classified Labels for the Same.

GAE

It is a Pytorch Library containing Simple and Fast implementations of Adversarial Attack Strategies using Pytorch. Cleverhans-Future and Advertorch can be referred for proper Robust implementations. GAE is easy to understand and only Processes a single image file at a time (as of now). Usage of the Following can be found on Attacks Tutorial on Imagenet.ipynb notebook.

Framework

- Python `Flask`-based server
	- Python backend allows access to Pytorch 
- Front-end using JQuery and Bootstrap