End-to-End Multimodal Data Augmentation and Adversarial Robustness Benchmark with AugLy for Images, Text, Audio, and PyTorch

This tutorial demonstrates how to use AugLy, a tool for modifying data across images, text, and audio, to test whether machine learning models remain reliable when their inputs are altered or attacked. The work covers practical techniques including creating synthetic datasets, applying various transformations while tracking their intensity, detecting copied images despite distortions, and evaluating text classifiers against adversarial manipulations and unicode obfuscation. The approach integrates these augmentation and robustness experiments directly with PyTorch's data handling systems, treating data modification both as a way to generate training examples and as a method to measure model robustness.