Towards Robustness Against Natural Language Adversarial Attacks

Explore the challenges and solutions in defending against natural language adversarial attacks in this seminar by VinAI researcher Luu Anh Tuan. Delve into various types of natural language attacks and their impact on deep neural networks in NLP systems. Examine recent defense strategies and their limitations. Learn about the innovative Adversarial Sparse Convex Combination (ASCC) method, which models the attack space as a convex hull and aligns better with discrete textual space. Discover how ASCC-defense generates worst-case perturbations and incorporates adversarial training to enhance robustness. Gain insights into a new class of defense techniques for NLP, including the potential of robustly trained word vectors to improve model resilience without additional defense measures.