The Role Of Machine Learning In Detecting Cyber Threats: A Comparative Review Of Techniques, Datasets, And Evaluation Challenges

Abstract

Traditional Systems That Rely On Signatures Or Rules Have Not Been Able To Keep Up With The Increasing Complexity, Frequency, And Effect Of Cyber Threats. There Has To Be A Dramatic Change To Threat Detection Systems That Use Machine Learning (ML) To Keep Up With The Ever-changing Threat Landscape. A Thorough Comparative Analysis Of ML's Function In Cyber Threat Detection Is Presented In This Review. We Aim To Thoroughly Analyze The Main ML Techniques, Identify The Domains Where They Are Applied, And Assess The Benchmark Datasets And Assessment Issues That Are Essential To This Field Of Study. Deep Learning (DL) Models Show Better Feature Extraction And High Accuracy (98-99%) In Complex Environments, According To A Comparative Study Of Supervised, Unsupervised, DL, And Hybrid/ensemble Methods. However, The Best Architectures For Maximizing Generalizability And Robustness Are Currently Hybrid And Ensemble Models. Problems With Out-of-date Content, Class Imbalance, And Real-world Representation Are Prevalent In Analyses Of Popular Datasets (e.g., CICIDS2017, NSL-KDD). Inadequate Dataset Quality, The Interpretability Problem Of "black Box" Models, And Susceptibility To Adversarial Attacks Are Some Of The Ongoing, Practical Obstacles That Limit ML's Efficacy, According To The Study's Conclusions. Establishing Adversarially Resilient, Explainable (XAI), And Data-quality-conscious ML Pipelines Is An Important Need For Future Research In Order To Guarantee The Deployment Of Reliable And Scalable Cyber Security Systems.

Keywords