The code in this repo pertains to a short comparative analysis I wrote titled "Effects of Gaussian and Adversarial Perturbations on a CNN's Performance". This paper can be found in the same repo in pdf form.
Convolutional Neural Networks (CNN) are powerful class of Deep Neural Networks (DNN) that are used mostly for image recognition and classification. They have proven themselves as one of the best models in modern ML and are widely accepted as the golden standard for image classification. Yet sometimes these networks are susceptible to certain types of data that can be harmful. In this paper we explore two distinct types of input data noise that can interfere with a CNN’s ability to perform correct classifications and to what degree these perturbations are harmful.
The dataset used in this paper is the MNIST database of handwritten digits. It is comprised of 60,000 training examples and 10,000 testing examples. Each sample is a 28 × 28 grey-scale image of a digit between 0 and 9.
Environment: - Python 3.7 - MacOS Catalina
Libraries: - TensorFlow - Numpy - Keras - Matplotlib