Adversarial Attacks in Some Recognition Models for Vietnamese Speech Classification

The last decade witnessed the tremendous development of Artificial Intelligence (AI). More and more real-life problems are involved with AI. Nonetheless, Machine Learning/Deep Learning models depend strongly on the data we feed in training/validation time. Training with noisy data was proved to cause harmful problems. This paper addresses adversarial attacks against recognition models for Vietnamese speech classification. The attacks are processes of creating audio samples that carry the original content; however, they distract the models with noises, which are hard to detect. Notably, we make attacks for different models that can be applied widely in practice to Vietnamese. This research is an essential first step in building measures to prevent such attacks in the future.