his research presents a method for classifying malicious and benign binary files using Convolutional Neural Networks (CNNs), transitioning from binary to multiclass classification. Three commonly used datasets were tested: EMBER, BODMAS, and MALIMG, with EMBER and BODMAS serving as training and testing sets for the base model. Data from these datasets is converted into image representations and analyzed by CNN models, achieving a high accuracy of 98%. A transfer learning model is then developed, incorporating knowledge from EMBER and BODMAS. This model reduces training time significantly and achieves 97% accuracy with just 5 epochs and a batch size of 25 across 25 malware family sets, averaging a perfect AUC of 1.00. This indicates perfect discrimination between positive and negative classes, with 100% correct predictions, underscoring the robustness of the method.