Hybrid Deep Learning Framework for Image Spam Detection Through Integrated Web Log Mining and Visual Feature Analytics

Abstract

Image-based spam remains a persistent threat because attackers embed textual message content into images to evade conventional text filters. This paper proposes a hybrid deep learning framework that fuses multi-channel visual feature analytics with web-log mining signals to improve detection accuracy and interpretability. The approach integrates convolutional neural networks (CNNs) for visual feature extraction, a transformer-based sequence module for web-log signal modeling, and a light-weight ensemble meta-classifier that combines the two modalities. We evaluate the method on public image-spam corpora and simulated web-log traces, compare it with state-of-the-art baselines, and report improvements in precision, recall and F1-score. Contributions include (i) a multimodal feature fusion architecture, (ii) a web-log feature taxonomy and mining pipeline for spam behavior profiling, (iii) explainability using SHAP, and (iv) an empirical evaluation with ablation and complexity analysis.