This paper introduces CGTdroid, a novel system for Android app risk assessment using cross-attention networks. CGTdroid detects privacy leaks and malicious behaviors by integrating static permissions, dynamic API calls, and network traffic features, encoded into 100-dimensional permission embeddings, 50-dimensional API sequences, and 20 network flow metrics. A cross-attention mechanism captures complex interdependencies across these features, outperforming traditional methods reliant on isolated or concatenated analyses. Utilizing LSposed hooks for runtime monitoring and an optimized Fastbot implementation, CGTdroid achieves comprehensive feature extraction within two minutes. Evaluated on 4,500 Android apps, including privacy-leaking and malicious samples, it attains 98.7% detection accuracy and 98.2% recall. CGTdroid’s scalable, interpretable, and robust framework effectively secures the Android ecosystem against evolving threats.