Advancing Research Knowledge.
Empowering Future Leaders.

Malware continues to evolve through obfuscation and packing, challenging static detectors and limiting the effectiveness of conventional vision-based CNNs that do not jointly capture multi-scale saliency and sequential dependencies encoded in malware images. This paper proposes ParaECA-LSTMNet — a purpose-driven integration of four parallel CNN branches to learn complementary spatial features, an Efficient Channel Attention (ECA) module to emphasize informative channels without added dimensionality, and an LSTM layer to model long-range dependencies before classification. Evaluated on the Malimg dataset (9,339 grayscale images across 25 malware families), images are resized to 224×224, normalized, and split using a stratified 70%/15%/15% train/validation/test protocol. ParaECA-LSTMNet attains 99.23% accuracy with Precision/Recall/F1 = 99.23%/99.20%/99.20%, outperforming EfficientNet-B0 (97.43%), VGG-16 + MobileNetV2 (94.57%), Inception-ResNet-V2 (93.23%), and Inception-V3 (92.52%). Confusion-matrix analysis confirms minimal inter-family confusion. The compact footprint makes the model suitable for deployment as a static pre-filter or in resource-constrained cybersecurity environments.

Accurate classification of peripheral blood cells is critical for diagnosing hematological disorders, yet manual microscopic analysis remains time-consuming and prone to inter-observer variability. This paper proposes an attention-guided dual-backbone fusion framework that combines the hierarchical shifted-window self-attention of the Swin Transformer with the self-supervised visual representations of DINOv2 to classify peripheral blood cells across eight morphological classes. The Swin Transformer branch captures multi-scale local cellular morphology through its shifted window attention mechanism, while the DINOv2 backbone contributes robust global visual embeddings learned via self-supervised distillation on large natural image corpora. A learnable attention-guided fusion module adaptively weights the contribution of each feature stream, enabling the model to emphasize the most discriminative cellular features at inference time. The framework is evaluated on the PBC dataset (17,092 images across 8 classes) and achieves state-of-the-art performance, outperforming single-backbone ViT, ResNet, and EfficientNet baselines. The model demonstrates strong generalization and clinical relevance for automated hematology diagnostic support systems.