Benchmark study finds 98.9% accuracy for concurrency-aware SDN intrusion detection

A concurrency-aware artificial intelligence framework reached up to 98.9% detection accuracy in benchmark experiments for real-time intrusion detection in software-defined networks, according to a peer-reviewed Scientific Reports article shared ahead of its final version. The system also delivered 2.1 times the throughput of a centralized intrusion-detection-system baseline and reduced average latency by approximately 57.5%, from 146 milliseconds to 62 milliseconds. The authors describe concurrency as a foundational architectural principle for scaling artificial-intelligence systems that recognize patterns in fast-moving data streams.

The study focuses on system architecture rather than proposing a new intrusion-detection model alone. Its concurrency-aware AI, or CA-AI, framework directly incorporates actor-based concurrency control, workflow-based load-aware scheduling, deep long short-term memory for temporal learning, agent-based parallel feature extraction, asynchronous inference, and monitoring-based feedback loops. Together, these elements coordinate the processing of network-traffic data as new flows arrive, addressing bottlenecks associated with inflexible scheduling, centralized learning pipelines, and limited fault tolerance.

Within this design, the deep LSTM component handles the temporal-learning component of intrusion detection, while actor-based execution and parallel feature extraction distribute concurrent processing work. Load-aware workflows schedule that work, and asynchronous inference allows the system to continue processing while other tasks are underway. Monitoring feedback connects operational conditions to the processing architecture. The supplied study abstract does not give enough detail to determine how each component was individually implemented or to quantify their separate contributions to the reported performance.

Researchers evaluated the framework using network-traffic samples from three established intrusion-detection benchmarks: NSL-KDD, CIC-IDS2017, and UNSW-NB15. The reported maximum accuracy across the datasets was 98.9%. Relative to the centralized IDS baseline, throughput increased by 2.1 times and average latency decreased by approximately 57.5%. The authors also report that CA-AI throughput increased by approximately 1.19 times when concurrency rose from 10,000 to 100,000 flows. These measurements come from controlled evaluations of recorded traffic samples, not a live deployment in an operational network or a test against a particular attack campaign.

The results support the authors’ narrower conclusion that concurrency-aware system design can improve the measured performance of AI-based intrusion detection under concurrent workloads. The 98.9% figure is a maximum reported across benchmarks, not a uniform result for every dataset, and the throughput and latency gains apply specifically to the authors’ experimental baseline and setup. The available abstract does not report uncertainty intervals, dataset-specific accuracy figures, or a detailed description of the centralized baseline, so the precision and generalizability of the comparisons cannot be assessed from the supplied evidence alone.

Subject of Research: Concurrency-aware AI for real-time intrusion detection in software-defined network benchmarks

Article Title: Concurrency aware artificial intelligence for real-time intrusion detection in software-defined networks

Article References: Kanimozhi, R., Padmavathi, V., & Latha, M. (2026). Concurrency aware artificial intelligence for real-time intrusion detection in software-defined networks. Scientific Reports. https://doi.org/10.1038/s41598-026-75362-x

Image Credits: AI Generated

DOI: 10.1038/s41598-026-75362-x

Keywords: intrusion detection, software-defined networks, concurrency, artificial intelligence, deep learning, long short-term memory, parallel computing, load-aware scheduling, network security, benchmark datasets, aware, artificial