Deep Neural Architectures for Predictive Threat Intelligence: Malware Detection via Optimized CNN–MLP Models

Abstract

The proliferation of sophisticated malware poses a critical threat to global cybersecurity, affecting individuals, enterprises, and critical national infrastructures. Traditional signature-based and heuristic detection methods frequently fail to identify zero-day attacks and obfuscated malicious code, motivating the development of robust, data-driven alternatives. This paper proposes a malware detection framework built upon a hybrid Convolutional Neural Network-Multi-Layer Perceptron (CNN-MLP) architecture that jointly captures local spatial correlations and global non-linear dependencies within malware samples. Binary executable files are first transformed into twodimensional grayscale images, after which stacked convolutional blocks automatically extract high-level structural features and textural patterns that are largely invisible to conventional analysis. These hierarchical feature representations are subsequently flattened and passed to a deep MLP classifier equipped with batch normalisation and dropout regularisation, which maps the extracted patterns to their respective malware families. This integrated design leverages the feature-learning capacity of CNNs together with the classification efficiency of MLPs, providing a dual-layer defence mechanism. Extensive experiments on the Microsoft BIG 2015 benchmark dataset comprising 21,741 samples across nine malware families demonstrate that the CNN-MLP fusion achieves a classification accuracy of 96.8% and a false discovery rate of 3.5%, outperforming standalone CNN, MLP, SVM, and ensemble baselines. The framework is particularly effective against packed and obfuscated malware, yielding a 50% relative reduction in false discovery rates compared to the best single-architecture baseline, underscoring its suitability for realworld deployment in automated threat-intelligence pipelines.