A Review of Attack and Defensive Research on Speech Adversarial Samples

Speech is an important medium of communication between people and smart devices, which not only carries semanti c information, but also contains additional information such as ge nder, age, and emotion. With the rapid development of deep lear ning technology, the performance of various types of speech proce ssing tasks has been significantly improved, which is widely used i n the fields of emotion recognition, identity verification and smart home control. However, deep neural network (DNN)-based speec h recognition systems have certain security risks, such as against sample attacks. By making minor modifications to the input sam ples, an attacker can spoof the deep learning model, leading to its prediction errors, which triggers potential security risks. Therefo re, it is particularly important to study the attack and defense me thods of speech adversarial samples. First, this paper introduces t he basic concept of speech adversarial sample, then systematically combs the classical speech adversarial sample attack methods in recent years, and introduces their defense strategies. Finally, the r esearch direction of speech adversarial sample attack and defense is discussed.