Evaluating Temporal and Structural Anomaly Detection Paradigms for DDoS Traffic
Yasmin Souza Lima 1, Rodrigo Moreira 1, Larissa F. Rodrigues Moreira 1 · 2026
Read Evaluating Temporal and Structural Anomaly Detection Paradigms for DDoS Traffic by Yasmin Souza Lima 1, Rodrigo Moreira 1, Larissa F. Rodrigues Moreira 1 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-16575.txt.
Introduction
Lima and colleagues test time versus shape features for DDoS anomaly detection in 5G networks. Most studies fix one format without checking which feature space matches the data. Their framework chooses between the two before training, using self-link and PCA variance checks. When probes prove unclear, a hybrid option stays reserved as a future fallback, not a proven branch. On two datasets with three tools, shape features matched or beat time ones. The gap widened as time ties weakened.
Catalog record
| CREATOR | Yasmin Souza Lima 1, Rodrigo Moreira 1, Larissa F. Rodrigues Moreira 1 |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 17719 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.16575
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Evaluating Temporal and Structural Anomaly Detection Paradigms for DDoS Traffic Yasmin Souza Lima 1, Rodrigo Moreira 1, Larissa F. Rodrigues Moreira 1 Abstract Unsupervised anomaly detection is widely used to detect Distributed Denial-of-Service (DDoS) attacks in cloud-native 5G networks, yet most studies assume a fixed traffic representation, either temporal or structural, without validating which feature space best matches the data. We propose a lightweight decision framework that prioritizes temporal or structural features before training, using two diagnostics: lag-1 autocorrelation of an aggregated flow signal and PCA cumulative explained variance. When the probes are inconclusive, the framework reserves a hybrid option as a future fallback rather than an empirically validated branch. Experiments on two statistically distinct datasets with Isolation Forest, One-Class SVM, and KMeans show that structural features consistently match or outperform temporal ones, with the performance gap widening as temporal dependence weakens. 1 Introduction Modern network infrastructures increasingly rely on cloud-native designs, microservice decomposition, and serverless execution models, generating large volumes of telemetry at every layer of the stack. As threat activity intensifies and AI ( AI )-enabled tooling accelerates adoption [ Gartner, Inc. 2026 ] , the pressing question is how to turn pervasive telemetry into a representation that makes malicious behavior separable before any detector is even selected. The main issue is not detector choice, it is how traffic is represented before training. Most prior work commits to either temporal features or structural projections such as PCA ( PCA ) and optimizes the detector around that choice without testing paradigm fit. Recent advances span attention based sequence models for 5G ( 5G ) telemetry, autoencoders paired with Isolation Forest and SHAP, and clustering variants, but they still do not provide a repeatable rule to choose between temporal and structural representations in new deployments [ Moreira et al. 2023 ] . This paper proposes a pre-training decision framework for unsupervised DDoS ( DDoS ) detection. It uses two probes, lag-1 autocorrelation of an aggregated flow signal and cumulative PCA explained variance, to prioritize temporal versus structural representations before model selection. The study evaluates only these two branches on two datasets with opposing structure, one temporally dependent and one high-dimensional 5G benchmark with weak temporal signal, using IF ( IF ), OCSVM ( OCSVM ), and KMeans. The hybrid branch is retained as a conceptual fallback for inconclusive probe combinations and is not empirically validated here. Within this scope, the results show that structural features consistently match or outperform temporal ones when autocorrelation is weak. Our contribution is therefore not a new detector, but a lightweight pre-training characterization heuristic that helps narrow the representation choice before unsupervised modelling. 2 Related Work Unsupervised anomaly detection for network traffic and cloud-native systems has been addressed from multiple angles, yet no prior study explicitly compares temporal and structural feature paradigms or provides a decision rule for choosing between them. In the 5G and O-RAN domain, deep sequence models with attention have been used to detect abnormal Network Function interactions in the 3GPP ( 3GPP ) Service Based Architecture [ Tan et al. 2025 ] , Continuous Time Markov Chain models have targeted signaling anomalies via digital twins and safe reinforcement learning [ Prince and Prabhavathi Neelakandan 2026 ] , and LSTM autoencoders deployed in the Near-RT RIC have triggered secure slicing against malicious User Equipment [ Moore et al. 2025 ] . For broader network traffic, IF has been extended with X-means clustering to improve detection rates [ Feng et al. 2022 ] , paired with autoencoders and SHAP-based explanations [ Carrera et al. 2022 ] , and combined with ConvLSTM for spatiotemporal modelling [ Kumar et al. 2025 ] . Complementary efforts include calibrated one-class classification under anomaly contamination [ Xu et al. 2024 ] , dilated-convolution variational autoencoders for multi-scale patterns in mobile counters [ González et al. 2024 ] , multi-modal fusion of logs and traces in microservice architectures [ Zuo et al. 2020 ] , and a research agenda for context-aware detection in serverless environments [ Nguyen et al. 2025 ] . Table 1 summarises the coverage of each study across ten dimensions; columns mark whether the work targets 5G /O-RAN, fuses multiple telemetry sources, employs deep learning or specific detectors (IF, OCSVM ), includes explainability or mitigation, and, crucially, whether it compares temporal and structural paradigms or offers a paradigm decision rule. Table 1 : Short Related Work Comparison. Paper 5G or O-RAN Multi-source Telemetry Deep Learning Uses IF Uses OCSVM XAI Mitigation Temporal vs Structural Paradigm decision rule Open Release [ Zuo et al. 2020 ] [ González et al. 2024 ] [ Nguyen et al. 2025 ] [ Prince and Prabhavathi Neelakandan 2026 ] [ Tan et al. 2025 ] [ Moore et al. 2025 ] [ Carrera et al. 2022 ] [ Kumar et al. 2025 ] [ Xu et al. 2024 ] [ Feng et al. 2022 ] Our Approach Contribution Positioning. Unlike prior detector-specific studies, this paper addresses the paradigm choice by comparing temporal and structural representations for unsupervised DDoS detection across statistically diverse datasets. Our conceptual contribution is a lightweight decision procedure for unseen deployments using two diagnostic probes: lag-1 autocorrelation and PCA cumulative explained variance. This framework bridges detector-centric research and deployment-time decision-making, particularly in 5G environments where optimal traffic representation fluctuates. 3 Methodology The pipeline consists of six sequential steps, illustrated in Figure 1 : 1 dataset selection and characterization, 2 preprocessing and cleaning, 3 feature engineering under two paradigms, 4 unsupervised detection, 5 evaluation against ground-truth labels, and 6 cross-paradigm comparison. Input, network traffic dataset Compute ACF at lag 1 on an aggregated flow signal Is lag-1 ACF high? Compute PCA cumulative explained variance \geq 95\% variance with \leq 5 PCs, ? Temporal paradigm One Class SVM on rolling features Isolation Forest in temporal space Hybrid paradigm Combine temporal and structural features Structural paradigm K Means on PCA space Isolation Forest on structural features No Yes No Yes Figure 1 : Lightweight framework for prioritizing temporal or structural representations. 3.1 Datasets Step 1 . Two labelled datasets with distinct traffic profiles were selected to evaluate whether detection paradigm effectiveness depends on the statistical nature of the traffic. CICDDoS2019. The CICDDoS2019 benchmark [ Sharafaldin et al. 2019 ] contains bidirectional flow records captured during simulated DDoS campaigns. Each record includes 80 features extracted by CICFlowMeter, covering packet counts, byte volumes, inter-arrival times, and flag distributions. The dataset includes eight reflection-based attack types, each stored as a separate CSV file. A stratified sample of 30 000 flows per file was drawn, yielding approximately 240 000 records with a heavily imbalanced class distribution (benign flows constitute less than 0.4%). 5GAD. The 5G Attack Detection dataset (5GAD) [ Coldwell et al. 2022 ] provides simulated traffic from a 5G network slicing testbed. Records are fixed-length feature vectors of 1 024 dimensions stored as NumPy arrays. Labels are binary ( normal / attack ), evenly distributed across 48 320 samples. The original tensor shape was flattened to a standard two-dimensional matrix prior to analysis. 3.2 Preprocessing Step 2 . All experiments ran on an Intel Core i7 workstation with 32 GB RAM using Python 3.12 and scikit-learn 1.4. Numeric precision was reduced to float32