DUAL-PHASE DEEP LEARNING FRAMEWORK FORSTEGWARE DETECTION USING CNN ANDVAE-CONDITIONED CGAN

Abstract

The increasing sophistication of stegware—malware concealed
within digital images—poses a major challenge to modern cybersecurity. This study develops and compares two novel deep learning frameworks to detect such hidden threats. The first employs a customized Convolutional Neural Network (CNN) that captures minute spatial and statistical distortions within image data to classify clean and infected samples. The second approach utilizes a Conditional Generative Adversarial Network (cGAN), where latent noise features extracted from a Variational Autoencoder (VAE) serve as conditional inputs, enabling improved representation learning and enhanced discrimination between benign and stegwareinfected images. Experiments were conducted on a domain-specific stegware dataset comprising 20,000 images generated using multiple steganographic
embedding schemes and realistic payloads. The CNN achieved an accuracy of 94.3%, F1 score of 0.94, and AUC of 0.96, effectively identifying embedded malware patterns. In contrast, the VAE-conditioned cGAN outperformed the CNN, achieving an accuracy of 98.2%, F1 score of 0.98, and AUC of 0.987, particularly excelling under complex embedding conditions. Paired statistical tests confirmed that these improvements were significant (p ¡ 0.005). The comparative results highlight the advantage of combining
generative modeling and adversarial learning for stegware detection, offering a robust, adaptive, and interpretable solution for securing image-based communication channels against covert cyber threats.