FedSecureXAI: Multimodal Federated Learning for Android Malware Detection with Byzantine-Resilient Aggregation, Differential Privacy, and Explainable AI

Abstract

AbstractFederated learning (FL) allows organizations to collaborate on Android malware detection without sharing raw samples. However, malicious client updates can affect the global model, while model updates can also expose information about private training data. We present FedSecureXAI, a multimodal federated framework for Android malware detection. Byzantine resilience is the primary security objective, while differential privacy (DP) and privacy-preserving explainability address information leakage. The framework combines visual bytecode features from Vision Transformer (ViT-B/16), behavioral system call representations, and static Android features. Local model updates are protected with DP. Malicious client updates are filtered using Krum-based screening and coordinate-wise median aggregation. The framework also uses a privacy-preserving Deep SHAP protocol for instance-level explanations. On the Drebin+Cuckoo benchmark with 20 clients and highly heterogeneous non-IID partitioning 0.1, FedSecureXAI achieves 92.4% accuracy and a macro-F1 score of 0.914. Under 25% Byzantine client participation, accuracy remains at 91.1%, compared with 62.6% for the undefended configuration. RDP accounting gives a total bound under the stated mechanisms and assumptions. Privacy experiments evaluate membership inference and gradient-based reconstruction, while XAI experiments assess the quality of Deep SHAP explanations against alternative methods. The framework also achieves 91.1% accuracy on CICMalDroid-2020, indicating consistent detection performance across the two evaluated Android malware datasets. Keywords: Federated learning, Android malware detection, Differential privacy, Byzantine resilience, Explainable AI, SHAP, Multimodal learning, Vision Transformer