Universal Targeted Adversarial Attacks Against mmWave-Based Human Activity Recognition

Artificial Intelligence (AI) has been the key driver in the rapid advancement of modern networking technologies, encompassing both wired and wireless communication systems. Among these, millimeter Wave (mmWave) technology stands out as a promising tool for high-speed communication and the establishment of strong network connections. In addition to its communication capabilities, recent studies have also showcased the potential of mmWave in fine-grained sensing applications, such as Human Activity Recognition (HAR). The state-of-the-art HAR systems predominantly rely on artificial intelligence and deep learning techniques. Despite their effectiveness and robust nature, these technologies remain susceptible to adversarial attacks specifically targeting deep learning models. There has been no comprehensive examination of mmWave-based HAR systems in the existing work. In this chapter, we will investigate both white-box and black-box adversarial attacks on these mmWave-based HAR systems. Our strategy encompasses the two prevalent mmWave-based HAR models—voxel-based and heatmap-based, thereby enhancing the reach of our attack. Our evaluations underscore the effectiveness of our designed attacks, further emphasizing their potential applicability in real-world scenarios.