Unsupervised autoencoder with IQR threshold detects 5G DoS/DDoS traffic in benchmark study

A computational study of the 5G Network Intrusion Detection Dataset found that a deep autoencoder trained only on benign traffic, combined with an adaptive interquartile-range threshold, identified 321,012 of 321,357 denial-of-service and distributed denial-of-service attack samples in a held-out evaluation set. The results were reported in a 2026 research article, with a precision of 97.66%, recall of 99.89%, F1-score of 98.79% and ROC-AUC of 97.87% across the full evaluation set of 382,401 flows.

The framework addresses a central problem in intrusion detection: signature-based systems are designed around known attack patterns, while supervised machine-learning systems require labelled examples of attacks. The study’s authors argue that obtaining and maintaining sufficient labelled attack data can be expensive and time-consuming in operational 5G environments. Their proposed autoencoder therefore received no attack labels during training and was intended to model normal traffic before identifying deviations through reconstruction error.

The researchers evaluated the approach on 5G-NIDD, a dataset collected from a live 5G testbed and assembled from ten per-attack CSV files. It contains 728,316 network-flow records with 112 raw features, spanning benign traffic and nine DoS/DDoS attack categories. The features covered network, transport and application-layer information, including packet counts, byte volumes, flow duration, timing statistics, TCP flags, protocol identifiers and IP-header fields.

Before modelling, the records underwent a cleaning and feature-engineering pipeline. The researchers removed labels, attack metadata, network identifiers, fully missing or near-constant columns, and high-cardinality port and IP-identification fields. Nine low-cardinality categorical features were one-hot encoded, while sparse missing values in three remaining columns were imputed with medians. The resulting feature matrix contained 101 variables per sample.

Thirteen heavily right-skewed features were transformed using log(1+x), and all features were then MinMax-scaled to the range from zero to one using values calculated exclusively from the training partition. The researchers reported that heavy-tailed variables, including a destination-load feature reaching 8.4 × 10^10, caused losses on the order of 10^10 when processed without logarithmic compression. In an ablation using the same architecture with robust scaling but no log transformation, the authors reported an F1-score below 0.18.

The autoencoder used a symmetric encoder-decoder architecture that compressed the 101 input features into a 16-dimensional latent representation before reconstructing them. It used batch normalization, dropout with a rate of 0.20, a sigmoid output layer, Adam optimization with an initial learning rate of 10^-3, early stopping and learning-rate reduction. Training stopped at epoch 73, and the reported final training and validation mean squared errors differed by approximately 2.7 × 10^-5.

Only the 406,959 benign samples were used for training, validation and benign testing, divided 70%, 15% and 15%, respectively. All 321,357 attack samples were reserved for evaluation. Each sample received an anomaly score based on its mean squared reconstruction error, and the researchers compared percentile, mean-plus-three-standard-deviation and adaptive IQR thresholds. Using identical autoencoder weights, the reported F1-score ranged from 0.465 to 0.988 according to the threshold selected. The IQR rule used the third quartile plus 1.5 times the interquartile range and produced a threshold of 1.15 × 10^-5.

At that operating point, the system generated 9,732 false positives among 61,044 benign test flows, corresponding to a false-positive rate of 15.94%. The results were aggregated across all nine attack categories, so the study did not establish performance for individual attacks or compare slower approaches such as Slowloris and Torshammer with high-volume floods. The authors noted that the high false-positive rate could create substantial alert volume in a high-throughput deployment and that a less sensitive threshold or secondary triage stage might be needed.

The findings should also not be interpreted as evidence of zero-day generalisation. The model and its threshold were derived and tested using splits of the same 5G-NIDD collection, and no independent, out-of-distribution dataset was used to validate the approach. The authors identified per-category evaluation, comparisons with Isolation Forest, One-Class SVM and supervised classifiers under identical preprocessing, latent-space analysis, explainability analysis and testing on independent benchmarks as future work. They also reported that two submitted reconstruction-error figures were inconsistent with the final preprocessed results and required regeneration, further limiting interpretation of the current artifact.

Subject of Research: Unsupervised DoS and DDoS intrusion detection in 5G network-flow data

Article Title: Adaptive IQR-based unsupervised intrusion detection for DoS/DDoS attacks in 5G networks using deep autoencoders

Article References: Shreyansh, K., & Pranav, P. (2026). Adaptive IQR-based unsupervised intrusion detection for DoS/DDoS attacks in 5G networks using deep autoencoders. Discover Artificial Intelligence, 6(1), Article 1447. https://doi.org/10.1007/s44163-026-02482-8

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02482-8

Keywords: 5G networks, intrusion detection, deep autoencoders, unsupervised learning, DoS attacks, DDoS attacks, IQR thresholding, network security, Adaptive, IQR-based, unsupervised, intrusion