Work

Deep learning-based DDoS detection with error-space fidelity learning for cross-distribution generalization

Year: 2026
Type: article
Abstract: Distributed Denial-of-Service (DDoS) attacks pose a significant challenge to intrusion detection systems because network traffic characteristics and attack behaviors can change across deployment environments. Conventional deep learning-based DDoS detection methods are commonly evaluated under closed-world assumptions, where training and testing data follow similar distributions, which may overestimate real-world generalization. To address this limitation, this study proposes an error-space fidelity learning framework for evaluating DDoS detector reliability under cross-distribution conditions. The proposed framework trains deep learning models on individual DDoS attack families and evaluates their ability to generalize to previously unseen attack distributions. In addition to conventional detection metrics, reconstruction or prediction-error distributions are modeled using kernel density estimation and compared using Jensen–Shannon divergence to derive a fidelity score. The framework is evaluated using the CIC-DDoS2019 dataset with Autoencoder, VGG, ResNet, EfficientNet, and MobileNet architectures under a strict cross-attack-family evaluation protocol. Experimental results demonstrate strong detection performance across unseen attack families, with PR-AUC values ranging from 0.950 to 0.999. However, the fidelity analysis shows that high classification performance does not always imply distributional compatibility. EfficientNet provides the best overall balance between detection performance and fidelity, while MobileNet achieves competitive performance with lower computational complexity. The results demonstrate that error-space fidelity provides an additional deployment-oriented measure for assessing the reliability of DDoS detection models under distribution shift.
Language: English
Open Access status: gold
Estimated APC paid: $1,520
OpenAlex ID: W7235859564
FWCI:
Cites: 37
Cited by:
Related to: 10