Malware analysis increasingly requires models that evaluate various behavioral evidence without the computational cost associated with many deep learning systems. Hierarchical memory structures, variational inference, and complex meta-learning pipelines are hard to scale and susceptible to poisoned input sets. Due to these limits, simpler, discriminative, and real-world systems are needed. This study unifies lightweight, highfidelity models to ease four-angle virus detection. By substituting deep hierarchical graph modules with dual shallow GCN encoders for control-flow and system-call graphs, the Hierarchical Graph Embedding Classifier (H-GEC) generates a compact joint embedding that resists structural obfuscations. The Segmented Time-Series Malware Detector (STMD) splits behavioral logs into windows, which are evaluated by efficient GRU encoders to detect delayed Intent malware without a variational temporal decoupling process. As a pragmatic alternative to counterfactual-based causal inference, the Causal Feature Filtering (CFF) suppresses spurious correlations and improves generalization through conditional Independence-driven feature re-weighting. For few-shot detection, the Multi-Modal Alignment Network (MMAN) aligns opcode, system-call, and network-flow representations utilizing contrastive learning and lightweight CNN, GRU, and MLP encoders. Lastly, Weighted Federated Averaging (WFA) weights client contributions based on validation performance, making decentralized training more resilient to biased local updates. These methods enable scalable, multi-modal, and tamper-resilient malware security systems by improving accuracy, interpretability, and deployment efficiency.