The proliferation of interconnected computing environments has expanded the attack surface for sophisticated cross-platform malware, which employs behavioral polymorphism, multi-device propagation, and coordinated strategies to evade conventional detection. Current deep learning defenses remain largely platform-specific, lack cross-device correlation, and struggle with resource constraints and adversarial evasion. To overcome these limitations, this paper proposes a Behavioral Cross-Device Malware Defense System (BCD-MDS) that unifies modality-specific deep learning models within a centralized intelligence framework. The system leverages Transformers for desktop API sequences, Temporal CNNs for mobile behavioral timelines and Graph Neural Networks for IoT communication graphs, enabling tailored analysis per device category. A meta-learning correlation module synthesizes multi-modal behavioral embeddings to identify coordinated threats across platforms, while an intelligent elimination mechanism provides automated, context-aware remediation. Evaluated on STRIDS (desktop), CIC-AndMal2017 (mobile), and IoT-23 (IoT) datasets, the system achieves detection accuracies of 95.2%, 95.6%, and 94.3%, respectively, outperforming existing single-platform baselines. The results validate the efficacy of a behavior-aware, multi-modal deep learning approach in achieving scalable, adaptive, and explainable cross-platform malware defense.