Advancing cybersecurity a comprehensive review of AI-driven detection techniques (journalofbigdata.springeropen.com)

A new open-access survey in Journal of Big Data compiles and critically evaluates more than 60 recent studies (past four years) on AI-driven cyber-attack detection, presenting a systematic framework to compare machine learning (ML), deep learning (DL), and metaheuristic approaches across anomaly detection, classification, feature reduction, and evaluation metrics. The review maps common attack classes (malware, network intrusions, spam, botnets, insider threats, APTs), catalogs public datasets used in experiments, and summarizes methods and performance in comparative tables—giving researchers a ready reference to current capabilities and gaps in automated threat detection. Technically, the paper highlights DL’s key advantage—automatic learning of complex, nonlinear features that improves detection of evasive and novel threats where manual feature engineering (typical of classic ML) falls short. It also emphasizes the complementary role of metaheuristic algorithms for feature selection and hyperparameter optimization. Crucially, the authors identify persistent weaknesses: dataset diversity/realism, model adaptability to drift, high false-positive rates, and evaluation standardization. Their main implication for the AI/ML community is actionable: prioritize robust, adaptable pipelines (continual learning, realistic benchmarks, hybrid DL+metaheuristic optimization) and regular model updates to keep pace with evolving adversaries.