The article discusses the development of VIC-KD, a technique for enhancing the robustness of keyword spotting (KWS) systems against adversarial attacks. KWS is a process of identifying predetermined words in audio streams, and is commonly used in voice assistant technologies. However, these systems face challenges due to hardware limitations and increasing adversarial attacks. VIC-KD is proposed as a solution, offering robust distillation for model compression and adversarial robustness. The method shows notable improvements over current robust distillation methods, increasing robust accuracy significantly.
Publication date: 25 Sep 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2309.12914
