Scholarlink

Discerning malware from trustworthy applications in Android has become imperative due to the increasing occurrence of fraud losses in this market area. To comply with General Data Protection Regulation, we propose an Android malware detector by exploiting federated learning to train a shared model, incorporating an additional layer of privacy preservation using a differential privacy aggregator. In this way, the trained model not only gathers all the characteristics of the models trained on locally preserved datasets but also ensures that the aggregation of these models is robust against sniffing attacks by malicious users. Specifically, we tested our model, built on top of the MobileNetV3 architecture, with default and differential privacy aggregators to analyze the properties of these techniques and their effect on the model’s performance and generalization capabilities. Additionally, to simulate a real-world scenario as closely as possible, we distributed the dataset into Independent and Identically Distributed and non-Independent and Identically Distributed settings, reduced the training parameters of the model, and compressed the size of the dataset to ensure compatibility with the majority of smartphones’ CPUs and RAM available in the today market. The results demonstrate that a trade-off between privacy and classification capabilities is possible with an accuracy of 84.1%.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Federated Learning-Based Android Malware Detector Through Differential Privacy

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Christian Peluso,
  • Giovanni Ciaramella,
  • Francesco Mercaldo,
  • Antonella Santone,
  • Fabio Martinelli

摘要

Discerning malware from trustworthy applications in Android has become imperative due to the increasing occurrence of fraud losses in this market area. To comply with General Data Protection Regulation, we propose an Android malware detector by exploiting federated learning to train a shared model, incorporating an additional layer of privacy preservation using a differential privacy aggregator. In this way, the trained model not only gathers all the characteristics of the models trained on locally preserved datasets but also ensures that the aggregation of these models is robust against sniffing attacks by malicious users. Specifically, we tested our model, built on top of the MobileNetV3 architecture, with default and differential privacy aggregators to analyze the properties of these techniques and their effect on the model’s performance and generalization capabilities. Additionally, to simulate a real-world scenario as closely as possible, we distributed the dataset into Independent and Identically Distributed and non-Independent and Identically Distributed settings, reduced the training parameters of the model, and compressed the size of the dataset to ensure compatibility with the majority of smartphones’ CPUs and RAM available in the today market. The results demonstrate that a trade-off between privacy and classification capabilities is possible with an accuracy of 84.1%.