Machine Learning Techniques for TCP/IP-Based Intrusion Detection Systems: A Systematic Literature Review

Authors

  • Mukaila Ishola Adediran Centre for Cyberspace Studies, Nasarawa State University, Keffi, Nigeria
  • S. A. Abdullahi Centre for Cyberspace Studies, Nasarawa State University, Keffi, Nigeria
  • Salma Muhammad Abidi Centre for Cyberspace Studies, Nasarawa State University, Keffi, Nigeria
  • Kenneth Y. Giwa Centre for Cyberspace Studies, Nasarawa State University, Keffi, Nigeria

Keywords:

Intrusion Detection System, , Machine learning, Deep Learning, TCP/IP Networks, Systematic Literature Review, Prisma 2020, Benchmark Evaluation, Minority Class Detection.

Abstract

The evidence base for Machine Learning-based intrusion detection is fragmented across incompatible benchmarks and unvalidated deployment claims. This PRISMA 2020 compliant systematic review synthesised 50 primary studies (2017 - 2026) from five electronic databases across four research questions on technique taxonomy, benchmark adequacy, comparative performance, and open challenges. Classical ML (random forest dominant, n = 23) and deep learning (LSTM/CNN dominant, n = 24) are co-dominant paradigms, ensemble methods yield at most 0.003 F1 macro gain over their strongest base classifier with no reliable diversity advantage on the principal benchmarks. Evaluation concentrates on three legacy datasets (NSL-KDD, CICIDS-2017, UNSW-NB15) with documented validity limitations with only three studies use the NF-v2 NetFlow family, the sole framework enabling controlled cross study comparison. Temporal interarrival features raise multiclass macro F1 by 19.8-27.3 percentage points over static flow features while binary gains remain below 2.2 points, exposing a systematic masking effect in aggregate accuracy reporting. Minority class failure is mechanistically rooted in feature poverty rather than sample scarcity. Eight convergent gaps are identified: adversarial robustness is untested in all 50 studies; concept drift is addressed in only one. Structural reform of the evaluation paradigm is required before ML-based IDS research can produce reproducible, comparable, and deployable results.

DOI: https://doi.org/10.5281/zenodo.23385324 

Downloads

Published

2026-10-15