Android Malware Detection using Federated Learning

This paper proposes a privacy-preserving Android malware detection framework based on Federated Learning (FL). We create a detection framework that employs a multimodal feature set that includes permissions, intents, and hardware specifications. Then, we utilized 5 simulated clients to train a lightweight fully connected neural network locally that emulates non-IID (independent and identically distributed) data distribution using the Flower FL framework. The framework includes a binary cross-entropy loss function and the Adam optimizer to achieve efficient training while preserving raw data on-device to protect user privacy. Our experimental results report high accuracy in malware detection while keeping privacy intact, which facilitates applicability to real-world mobile scenarios.