Online Gambling Detection Systems: A Systematic Review

Background Illegal online gambling increasingly exploits websites, mobile applications, social media, messaging services, and decentralized financial infrastructure, and operators actively evade automated filtering. This review synthesizes how online gambling detection systems (OGDS) are applied, integrated, and evaluated, and which limitations constrain their deployment. Methods We searched IEEE Xplore, ACM Digital Library, ScienceDirect, and Scopus on 9 January 2026. Peer-reviewed English-language journal and conference studies in computer science published from 2021 to 2025 and reporting measurable evaluation metrics were eligible. We appraised methodological quality using a 10-item checklist and quantified inter-reviewer agreement with Cohen’s kappa. We synthesized evidence through frequency analysis and thematic synthesis across four research questions. Results Of 1,866 records identified, we included 39 studies (kappa = 0.83). Website detection dominated the corpus (15 studies), followed by harm detection (9), social media and spam messages (6), crypto and blockchain (4), mobile applications (3), and network infrastructure (2). Researchers identified six architectural integration patterns: multimodal feature fusion, graph-structural integration, hybrid ensemble learning, transformer-based semantic systems, harm systems, and rule-based operational systems. Multimodal fusion and graph-based models reported the highest detection scores against content obfuscation, whereas lightweight rule-based tools were favored for low-resource infrastructure. Evidence quality was uneven: 12 studies lacked a baseline comparison, 27 did not release data or code, and only six reported computational cost; differing datasets and protocols preclude direct performance comparison. Four recurrent limitations emerged: data quality and temporal drift; adversarial evasion and multimodal obfuscation; semantic ambiguity of localized slang and blockchain anonymity; and computational constraints. Conclusions Research priorities include adaptive recalibration, lightweight edge deployment, deep semantic entity extraction, explainable models for legal admissibility, and Web3 forensics. Shared, longitudinal, and cross-regional benchmarks are needed before performance claims can be generalized beyond individual study settings.