Adversarial attacks and graph regularization

Demonstration of adversarial attack on trained deep learning models using deceptive data,and use of graph regularization to build models that are robust to such attacks.

---

Notebooks

• fgsm-adversarial-attack
• gnn-graph-regularization

---

Fast gradient sign method

The fast gradient sign method works by using the gradients of the neural network to create an adversarial example. For an input image, the method uses the gradients of the loss with respect to the input image to create a new image that maximises the loss. This new image is called the adversarial image. This can be summarised using the following expression:

  adv_x = x + ϵ*sign(∇J(θ,x,y))

Addition of adversarial pattern

Misclassification on addition of adversarial pattern

Key observation : The confidence of the image classified as 0 decreases and goes as low as 0.02% on addition of pertubation, while the image still remains visually similar to a zero.

---

---

Graph regularization

Dataset

Access the dataset here.

This dataset contains four types of shapes namely, category 1, category 2, category 3 and category 4.

Improvement of validation accuracy and robustness graph regularized CNN model:

• With the use of a graph regularized deep neural network instead of the conventional base CNN model, a significant improvement in learning performance and validation accuracy can be achieved leading to a more robust CNN model which is fairly immune to adversarial attacks carried out with the help of techniques such as FGSM as explained above.
• The tensorflow neural structured learning module has been harnessed for synthesizing graph data structure to keep track of training examples that are similar and the degree of similarity that they have among themselves, which has then been fed to the neural network along with the training data to achieve graph regularization.

Comparison of performances of both models

• Key observation: The performance of graph regularized model is better on validation set as Compared to base CNN model, as it is more robust to deceptive samples.

---

Tools and technologies

Deep learning framework: tensorflow, nsl

Other tools and frameworks: scikit-learn, streamlit, dvc, jupyter notebook

---

Hosting

The web application has been hosted here.

Run Locally

Clone the project

  git clone <link-to-this-repo>

Go to the project directory

  cd gnn-adversarial-regularization

Install dependencies

  pip install -r requirements.txt

Start the server

  streamlit run src/app.py