/Trustworthy-AI-Projects

Project Overview

This repository details a two-part project aimed at improving model generalization and robustness in machine learning applications. Below is an overview of each part and the techniques employed.

Part 1: Enhancing Model Generalization

The first segment of the project focuses on training a model that achieves acceptable accuracy on both the SVHN and MNIST datasets.

Phase 1

  • Training Data: Used the SVHN training set.
  • Evaluation Data: Tested on both the SVHN and MNIST test sets.
  • Methods Employed:
    • Pre-trained ResNet Model: Leveraged for its robust feature-extracting capabilities.
    • Label Smoothing: Implemented in the loss function to help generalize better to unseen data.
    • Data Augmentation: To increase the diversity of the training data.

Phase 2

  • Reverse Training Approach:
    • Training Data: Switched to the MNIST training set.
    • Testing Data: Evaluated on both SVHN and MNIST test sets.
    • Fine-Tuning: Adjusted model parameters to enhance accuracy and generalization.

Part 2: Improving Robustness Against FGSM Attacks

The second part of the project addresses the model’s defense against FGSM attacks, using the CIFAR10 dataset.

  • Adversarial Data Creation: Developed a method to generate adversarial examples.

  • Training Conditions:

    1. Cross-Entropy Loss: Standard training to establish a baseline.
    2. Adversarial Training: Included domain adaptation techniques with adversarial examples.
    3. Circle Loss: Implemented according to the method described in this Circle Loss Paper.

  • Visualization: Employed UMAP plots for better visualization and analysis of model performance under various conditions.