As the frequency of malware attacks rises, affecting numerous users, businesses, and government entities, the importance of malware detection research has intensified. Contemporary malware variants often adopt evasion techniques such as polymorphism and metamorphism to swiftly modify their behaviors and produce numerous iterations. These new threats mainly consist of adaptations of existing malware, which has led to an increased use of machine learning algorithms for detailed analysis. However, these techniques can be time-consuming, requiring significant feature engineering, learning, and representation. To address this research gap, this study examines various deep learning architectures for malware detection, classification, and categorization using a specific dataset. It also mitigates dataset bias in the experimental process by applying distinct dataset splits to train and test the model independently across different timescales. The primary innovation is the development of a novel deep learning framework aimed at establishing an effective model for zero-day malware detection. The Inception-ResNet-v2 model is employed for analyzing malware families. A detailed comparison indicates that this proposed deep learning framework outperforms traditional machine learning algorithms.