Confidential · Professional · Structured · Research & publication support worldwide
Home / Blog / AI-Based Network Intrusion Detection: How to Plan a Publishable Thesis Study
Cybersecurity • AI

AI-Based Network Intrusion Detection: How to Plan a Publishable Thesis Study

A practical guide to AI-driven intrusion detection research: defining the task, choosing datasets, avoiding leakage, baselines, metrics and publishable experiments.

Network intrusion detection is one of the most active thesis areas in cybersecurity. The idea is easy to state: a system watches network traffic and raises an alert when something looks like an attack, such as a port scan, a denial-of-service flood, brute-force login attempts or botnet traffic. Machine learning and deep learning promise to detect attacks that fixed signature rules miss. But a thesis that simply trains a classifier on a public dataset and reports a very high accuracy rarely survives a serious examiner or journal reviewer. This guide explains how to plan an AI-driven intrusion detection system (IDS) study that is technically honest and publishable.

1. Define the detection problem precisely

"Intrusion detection" can mean several different tasks. Choose one primary task and state it clearly at the start of your proposal:

  • Binary detection: is a flow or packet window normal or malicious?
  • Multi-class classification: which attack family is it, for example DoS, probe, brute force, web attack or botnet?
  • Anomaly detection: the model learns only normal behaviour and flags deviations. This is closer to detecting unseen (zero-day) attacks, but usually produces more false alarms.

Also decide the unit of analysis: individual packets, bidirectional flows, or time windows of host or IoT device activity. This choice determines your features, your labels and which baselines are fair.

2. Choose datasets you understand

Public datasets make your work reproducible and comparable. Commonly used examples include NSL-KDD (a cleaned version of the older KDD Cup 99 data), UNSW-NB15, CIC-IDS2017 and CSE-CIC-IDS2018 from the Canadian Institute for Cybersecurity, and IoT-focused sets such as ToN_IoT and CIC-IoT2023. Each has known strengths and weaknesses. NSL-KDD is old and does not reflect modern traffic, so it is weak as the only dataset in a new thesis. Several researchers have also published analyses of labelling and feature-extraction problems in popular flow datasets, so read the documentation and any published critiques before you trust the labels.

A stronger design uses at least two datasets, ideally one recent general-purpose set and one domain-specific set (for example IoT or SDN traffic). If you generate your own traffic in a lab or simulator, describe the topology, tools, attack scripts and labelling process in full.

3. Avoid the leakage traps that inflate results

Many IDS papers report near-perfect scores that disappear in realistic testing. The usual causes are:

  • Identifier features: source and destination IP addresses, ports or timestamps can let a model memorise which host was the attacker in that capture. Remove or justify them.
  • Random splitting of correlated flows: flows from the same attack session end up in both training and test sets. Prefer splitting by time (train on earlier days, test on later days) or by capture.
  • Preprocessing on the whole dataset: fit scalers, encoders, feature selection and oversampling methods such as SMOTE on the training split only.
  • Duplicates: remove exact duplicate records before splitting, or report that you checked for them.

4. Build models in a defensible order

  1. Classical baselines: Random Forest, XGBoost or LightGBM on flow features. On tabular flow data these are often very strong, and examiners expect to see them.
  2. Deep baselines: a multilayer perceptron, a 1D-CNN, and an LSTM or GRU over sequences of flows or packets.
  3. Your proposed approach: for example an autoencoder or other anomaly detector for unseen attacks, a Transformer over flow sequences, a graph neural network over host communication graphs, a federated IDS for distributed IoT devices, or a lightweight model for edge deployment.

If a tuned tree-based model matches your deep model, report it honestly and argue for your method on other grounds such as generalisation, latency or privacy, if your experiments support that.

5. Report metrics that reflect real operations

Attack traffic is usually imbalanced, so accuracy alone is misleading. Report precision, recall and macro F1-score, per-class results with a confusion matrix, and the false positive rate, because in a real security operations setting a small false alarm rate on large traffic volumes still means many alerts. For anomaly detection, add the area under the precision-recall curve. Run multiple random seeds and report mean and standard deviation.

6. Add the experiments that make it publishable

  • Cross-dataset or unseen-attack tests: train on one dataset and test on another with aligned features, or hold out an entire attack family during training.
  • Ablation study: show what each component of your model contributes.
  • Explainability: use SHAP or similar methods to show which features drive alerts, and check whether they make sense to a network engineer.
  • Efficiency: report model size, training time and inference time per flow, especially for IoT or edge scenarios.
  • Robustness: where relevant, test sensitivity to noise, concept drift over time, or simple adversarial perturbations.

7. Turn it into a research question

A focused question gives your thesis direction. For example: "Can a lightweight anomaly detector trained only on normal IoT traffic detect attack families not seen during training, with an acceptable false positive rate?" or "Does modelling host communication as a graph improve detection of botnet traffic compared with flow-level classifiers under a time-based split?" A direction like this could be novel, but novelty is only established after a careful literature comparison, a full implementation and fair baseline experiments.

Quick checklist before you write

  • Task, unit of analysis and threat model clearly stated.
  • At least one recent dataset, documented and cleaned, with leakage-free splits.
  • Strong classical and deep baselines under identical conditions.
  • Macro F1, per-class results and false positive rate over several seeds.
  • Generalisation, ablation and explainability analyses.

Working on intrusion detection or AI for cybersecurity?

If you have your own research idea in network intrusion detection, anomaly detection or AI-driven cybersecurity, the FHJCIS team can work with you on literature review, experiment design, Python implementation, data analysis and writing, and help you prepare the work for journal submission. Authorship and credit follow actual contribution. WhatsApp: +92 331 2663342 or +92 312 6243342.

Research SupportData AnalysisWhatsApp usContact

Related

Chat on WhatsApp