Adversarial Attacks on Voice Recognition Based on Hyper Dimensional Computing

Recently, there is a great demand for experimenting with Artificial Intelligence (AI) algorithms on the Internet of Things (IoT) devices that have only limited computing or transmission resources. Hyper-Dimensional Computing (HDC), which can effectively run on low-cost CPUs, is one of the solutions. However, since the AI algorithms are proved to be vulnerable to Adversarial Examples (AE) in recent research, it is then important to investigate the same security issues on other intelligent algorithms such as HDC. In our paper, motivated by the AE attacks for AI algorithms, we propose an attack measured based on the Differential Evolution (DE), which does not rely on the gradient. By attacking the VoiceHD model in the Isolet dataset, we prove that HDC is also vulnerable to AEs. In our experimentation, we can launch non-targeted attacks on the VoiceHD with the highest 85.7% success rate.