LightARM-Net: Accelerating Sparse-Attention Tabular Model for Android Malware Detection

Android malware continues to pose a major security risk as attackers develop increasingly sophisticated and evasive techniques. Although deep learning methods have shown strong potential for malware detection, existing models often suffer from high computational cost and inefficient attention mechanisms. This is the limitation of their use in real-time security systems. In particular, the Adaptive Relation Modeling Network (ARM-Net) relies on a bisection-based α-Entmax solver when α > 1, which introduces substantial computational overhead. To address this limitation, this paper presents LightARM-Net, a fast and efficient variant of the ARM-Net that leverage a Triton-optimized α-Entmax solver based on a hybrid Halley-bisection algorithm. Experiments on two benchmark datasets; TUANDROMD and KronoDroid (binary and multi-class), show that LightARM-Net consistently outperforms state-of-the-art deep tabular and classical machine learning models. The proposed model achieves up to a 2.4% improvement in F1-score while reducing the time per epoch by up to 4.3×, total training time (under early stopping) by up to 3.5×, and inference throughput by up to 8.5× compared with the original ARM-Net. These results indicate that LightARM-Net provides an effective and practical solution for efficient, real-time Android malware detection.