AI-Based Adversarial Method in Android

In recent years, the app of machine learning (ML) in Android security has attracted widespread attention, and its research enthusiasm continues to rise. Many scholars focus on exploring the potential and app of machine learning algorithms in Android security. This chapter will focus on the integration of machine learning technologies in the Android security field. In Android security, adversarial examples are an important research direction. Adversarial examples are generated by making slight modifications to the original sample so that it can fool the machine learning model. These modifications are usually within a range that is imperceptible to humans but can have a significant impact on the model. Research on adversarial examples helps to understand the vulnerabilities of machine learning models and propose corresponding defense strategies. Regarding adversarial examples, researchers have proposed different classification methods and attack techniques. Among them, black-box and white-box attacks are two common perspectives. A black-box attack means that the attacker can only infer the behavior of the model through its input and output, without direct access to the specific structure and parameters of the model. A white-box attack means that the attacker can fully access and understand the internal information of the model. In these two attack methods, researchers have proposed various adversarial methods and techniques. In order to conduct in-depth research on black-box and white-box confrontation, many typical works have been proposed. These works include different methods and technologies, such as transfer learning, meta-adversarial training, zero-order optimization, and query-based adversarial attacks. By introducing these works in detail, we can deeply understand the app of machine learning in Android security and related technologies. First, we will introduce adversarial examples and their principles. Second, we will introduce common adversarial methods and typical works from the perspective of black-box and white-box attacks.