Optimizing Online Gambling Site Detection via XLM-RoBERTa and ResNet34-Based Early Fusion

Abstract

The illegal online gambling ecosystem in Indonesia has evolved into a persistent cyber threat, driven by the adaptability of threat actors in manipulating content and network infrastructure. Conventional detection methods relying on domain reputation-based blocking (blacklists) and keyword matching now face systemic failure due to sophisticated evasion techniques such as domain hopping, SEO manipulation, and content cloaking. This study proposes an automated detection framework based on Multimodal Deep Learning that simultaneously integrates semantic, visual, and infrastructure metadata analysis. We employ an Early Fusion strategy by constructing a 1,284-dimensional combined feature vector, consisting of 768 dimensions of text embeddings from the XLM-RoBERTa model, 512 dimensions of global visual features from the ResNet34 architecture, and 4 hybrid technical metadata features. To ensure high-quality ground truth and address previous transparency concerns, this approach is evaluated using a balanced dataset of 3,546 sites constructed via active crawling using the Playwright framework. The data was rigorously verified by a panel of five security experts using a majority voting scheme, achieving a Fleiss’ Kappa agreement score of 0.87, which indicates almost perfect consensus. Experimental results demonstrate that the Early Fusion model with Random Forest classification achieved an F1-Score of 95.52%, significantly outperforming the Late Fusion strategy (91.62%) and other unimodal approaches. Furthermore, this study empirically confirms the phenomenon of adversarial adaptation, revealing that traditional metadata features contribute only 0.62% to the classification decision. These findings underscore the urgency of a paradigm shift towards Deep Content Inspection to modernize national cybersecurity filtering infrastructure.