The rapid evolution of malware including polymorphic and metamorphic variants limits the effectiveness of traditional signature and heuristic based detection methods. Since Windows Portable Executable (PE) files remain a primary attack vector advanced techniques are needed for reliable classification. This paper presents a malware detection framework that integrates Function Call Graphs (FCGs), Graph2Vec embeddings and Graph Attention Networks (GAT) to capture structural dependencies and improve detection accuracy. The dataset collected from MalwareBazaar and the Dike Dataset contains 1,885 balanced benign and malicious samples. Our approach involves Control Flow Graph (CFG) recovery, FCG generation using Angr and feature extraction with Graph2Vec, Node2Vec and Structure2Vec embeddings. We evaluate multiple models including XGBoost, Convolutional neural Network (CNN), Graph Convolutional Networks (GCN) and Graph Isomorphism Networks (GIN) with GAT. Results show that graph neural networks consistently outperform traditional models with Graph2Vec + GAT achieving the best performance 98.67% accuracy, 98.73% F1-score and 98.38% AUC. These results highlight the value of structural feature modeling and attention mechanisms for malware detection. While limited by dataset size the framework demonstrates strong potential for detecting emerging Windows malware and underlines the promise of graph based deep learning in cybersecurity.