With 249 million cases and 608,000 fatalities recorded in 2022, malaria is one of the major worldwide health concerns, particularly in areas with low resources. In this paper, we propose a custom convolutional neural network (CNN) with an integrated attention mechanism to inspect malaria from blood smear images. To improve model robustness, we combined three publicly available datasets from the NIH and Kaggle. The proposed model achieved 98.20% accuracy, 97.85% precision, 98.55% recall, and 98.20% F1-score, outperforming conventional di agnostic methods. In addition, we conduct comparative analyses using two transfer learning models, ResNet50 and DenseNet. ResNet50 attained 95.06% precision, 95.44% recall, with 95.05% F1-score, while DenseNet achieved a pre cision of 87.96%, recall of 88.33%, and F1-score of 87.90%. For interpretability, Grad-CAMandsaliency map visualizations highlighted key image regions, with saliency maps offering finer pixel-level localization. These results highlight the potential of our attention-based CNN as a feasible, interpretable diagnostic tool for malaria, particularly in low-resource settings.   (Ghosh et al., 2025)