Hybrid deep learning models for detecting user datagram protocol-based distributed denial of service attacks in internet of things networks

Abstract

Distributed denial of service (DDoS) attacks, particularly those that employ the User Datagram Protocol (UDP), are now a threat to Internet of Things (IoT) networks. Traditional intrusion detection systems (IDS) usually fail to detect these threats in real time because IoT devices are small and have minimal computing power. In order to address this problem, this study proposes a set of hybrid deep learning models that include dimensionality reduction and classification techniques for improved efficacy and precision in detecting UDP-based DDoS attacks. Four hybrid frameworks were developed and evaluated using the CIC-IoT2023 dataset: PCA–SVM, PCA–RNN, Autoencoder–SVM, and Autoencoder–RNN. Each model combines statistical feature extraction with either margin-based classification or temporal learning to balance computing cost and precision. The PCA–RNN and Autoencoder–RNN models demonstrated the highest detection accuracy, surpassing 99%, along with minimal prediction latency and good generalization, according to the results. In contrast, PCA-SVM showed faster inference but comparatively lower accuracy, which makes it suitable for thin-client IoT edge applications. The hybrid models also fared better in terms of scalability and performance when compared to CNN, LSTM, and Transformer-based IDS frameworks. In IoT scenarios with limited resources, the results confirm that feature reduction and sequential learning can be used to enhance IDS performance. The foundation for future research into attention-driven and adaptive intrusion detection methods in developing IoT infrastructures is laid by this work, which also provides a helpful avenue for developing hybrid IDS solutions that are lightweight, scalable, and real-time.