FedDroidMeter: A Privacy Risk Evaluator for FL-Based Android Malware Classification Systems.

Area of Science:

• Cybersecurity
• Machine Learning
• Privacy-Preserving Technologies

Background:

• Traditional Android malware classifiers collect sensitive user data, posing privacy risks.
• Federated learning (FL) offers a privacy-preserving alternative but remains vulnerable to indirect privacy inferences.
• Existing privacy evaluation methods lack a comprehensive approach for FL-based malware detection.

Purpose of the Study:

• To propose FedDroidMeter, a privacy risk evaluation framework for FL-based Android malware classifiers.
• To measure the privacy risks associated with sensitive information disclosure in these classifiers.
• To provide a systematic approach for assessing and comparing privacy risks across different FL configurations.

Main Methods:

• Developed FedDroidMeter based on normalized mutual information to quantify privacy risks.
• Designed the framework to be independent of specific attack models and capabilities.
• Conducted numerical assessments using the Androzoo dataset and baseline FL classifiers.

Main Results:

• FedDroidMeter effectively measures privacy risks in FL-based Android malware classifiers.
• The framework enables equal comparison of privacy risks across different models, FL settings, and privacy parameters.
• Preliminary studies explored the fundamental laws governing privacy risk within these classifiers.

Conclusions:

• FedDroidMeter provides a crucial systematic framework for evaluating privacy risks in FL-based malware classifiers.
• The findings highlight the importance of privacy risk assessment in developing secure FL systems.
• This research offers a theoretical basis and practical experience for developing targeted privacy defense methods.