Research on Deep Learning-Based Android Malware Detection Systems

To combat evolving Android malware, this paper proposed a lightweight deep learning detection system leveraging Drebin, AndroZoo Lite, and Canadian Institute for Cybersecurity MalDroid 2020 datasets. The approach fused static (permissions, call graphs) and dynamic (behavior logs) features via a hybrid convolutional neural network-Transformer model with bidirectional information flow for end-to-end training. To meet mobile device constraints, joint optimization through network pruning, quantization, and attention distillation was applied. Evaluated via five-fold cross-validation, the method outperformed baselines (support vector machine, long short-term memory, BERTroid (BERT-based Android Malware Detection Model), convolutional neural network-Vision Transformer) in precision, recall, F1, area under the curve, and inference latency, achieving high accuracy with low delay. It remains robust against polymorphic and obfuscated variants. Error analysis reveals the critical impact of feature fusion weights on decision-making, offering insights for real-time mobile threat defense.