Malware detection remains challenging because of privacy constraints, heterogeneous data distributions, and vulnerability to adversarial attacks in distributed environments. We proposed a federated adversarial learning framework for robust, privacy-preserving malware classification. The framework introduces a two-way federated strategy, in which two client groups train vision transformers using different patch sizes to capture fine- and coarse-grained malware features. A hierarchical selective aggregation mechanism enhances the global model by weighting trustworthy client contributions and improving resilience to stragglers and heterogeneous data, while integrated adversarial training mitigates model poisoning and data leakage. Experiments conducted on the Malevis, BIG2015, and Malimg benchmarks demonstrated strong scalability, achieving high accuracy with four clients, and minimal performance degradation at eight and 12 clients. The proposed approach further outperformed standard and adversarial attention-based vision transformers as well as state-of-the-art federated approaches, delivering a scalable, secure, and interpretable solution for modern malware detection.