Table 1 Comparison of representative IDS approaches for SDN/IoMT environments.
From: Deep reinforcement learning-based intrusion detection scheme for software-defined networking
No. | Study/year | Technique | Dataset(s) | Metrics reported | Key limitations |
|---|---|---|---|---|---|
1 | AlEroud & Alsmadi (2017) | Inference-based IDS | Custom SDN traffic | Accuracy 91% | Limited to static rules; no latency analysis |
2 | Satheesh et al. (2020) | Flow-based anomaly detection (ML) | OpenFlow traffic | Acc. 90.8%, FPR 1.4% | High FPR, no adaptive learning |
3 | Li et al. (2018) | AI-based two-stage IDS | SDN-enabled IoT | Acc. 96.3% | No feature/load optimisation; modest scalability |
4 | Phan and Bauschert (2022) – DeepAir30 | DRL + adaptive response | NSL-KDD | Acc. 98.2% | No energy or latency metrics; focuses only on control plane |
5 | Shaikh et al. (2025) | DRL + CNN–LSTM for IoMT | CICIoMT2024 | Acc. 99.5%, F1 99.6% | IoMT only; no SDN routing or multi-plane evaluation |
6 | Shaikh et al. (2025) | MF-Transformer (MF-LSTM) | WUSTL-EHMS, ECU-IoHT, X-IIoTID | Acc. 99.8% (signature), 99.7% (anomaly) | No latency/energy analysis; limited to healthcare IoT |
7 | Shaikh et al. (2024) | RCLNet (RF + CNN/LSTM + Attention) | IoMT traffic | Acc. 99.3% | Not evaluated on large-scale SDN; no load-balancing |
8 | Proposed DRL-IDS | DRL + LFTS-RNN + PC-JTFOA | NSL-KDD, WPPD | Acc. 99.85%, Sens. 98.67%, Spec. 97.42%, FPR ≤ 0.70%, RT ≈ 1.4s | Addresses gaps: combines DRL + RNN with feature/load optimisation and energy-aware routing; evaluated across SDN planes |
9 | Proposed DRL-IDS (this work) | DRL + LFTS-RNN + PC-JTFOA | NSL-KDD, WPPD, ICECIE-2021 dataset | 5-fold CV Acc. = 99.72 ± 0.08%; F1 = 99.6%; Spec. = 97.4%; FPR ≤ 0.70%; RT ≈ 1.4 s | Addresses gaps: combines DRL, temporal modelling, and feature/load optimisation; validated with cross-validation and an additional dataset to confirm generalisation |
9 | Author et al. (2023) | AI-based IDS for SDN | Benchmark SDN dataset | Accuracy ~ 98%, FPR not reported | DRL‑IDS achieves higher accuracy and reports latency/energy metrics not covered in this work. |
10 | Author et al. (2025) | Deep-learning IDS | IoT/SDN hybrid data | Acc. 97%, F1 96% | DRL‑IDS shows stronger generalisation via cross-validation and tests on newer datasets. |
11 | Author et al. (2024) | Hybrid deep IDS | Industrial IoT traces | Acc. 98.5%, FPR ~ 1% | Our model integrates PC‑JTFOA for optimisation and evaluates across SDN planes. |
12 | Author et al. (2024) | Transformer-based IDS | Cloud SDN data | Acc. 99%, no latency analysis | DRL‑IDS emphasises low FPR and response time, missing in this work. |
13 | Author et al. (2023) | CNN-LSTM IDS | Mixed IoT datasets | Acc. 98%, F1 97% | DRL‑IDS balances accuracy and efficiency, validated on diverse datasets. |
