Assessing the Effect of Model Poisoning Attacks on Federated Learning in Android Malware Detection

Abstract

Android devices are central to our daily lives, which leads to an increase in mobile security threats. Attackers try to exploit vulnerabilities and steal personal information from the installed applications on these devices. Because of their widespread usage, these devices are the prime targets of cyber attacks. To get rid of this, Android malware detection has become increasingly significant. Federated learning, which is a decentralized machine learning approach, has been utilized to improve the privacy of sensitive user data. However, the integration of federated learning also introduces a vulnerability to model poisoning attacks, where adversaries deliberately bias the learning process of the model to impair the performance metrics. This paper presents a comprehensive assessment of the effect of model poisoning attacks on federated learning systems deployed for Android malware detection. We also explain an exhaustive feature selection methodology that employs both static and dynamic features of Android applications and created a novel dataset. We focus on incorporating recent malware samples while creating the dataset to make the model robust and adaptable to new malware. Furthermore, we quantify the degradation in model accuracy and reliability following a model poisoning attack scenario through a series of experiments. Additionally, we explore the defense mechanisms to mitigate the model poisoning attacks based on recent studies. 2024 ACM.