A Deployment-Oriented and Resource-Efficient Neuro-Symbolic Framework for Explainable DDoS Detection in Operational Technology Networks
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: We're looking at the paper: "A Deployment-Oriented and Resource-Efficient Neuro-Symbolic Framework for Explainable DDoS Detection in Operational Technology Networks".
Paper summary: Nadia: We're diving into this paper now called "A Deployment-Oriented and Resource-Efficient Neuro-Symbolic Framework for Explainable DDoS Detection in Operational Technology Networks". The main thesis here is that we need a framework that works well in operational technology networks, which are really constrained by computation, memory, and latency.
Elias: So the paper proposes fusing a compact gated recurrent unit with an interpretable decision tree to tackle this problem by achieving explainable DDoS detection in these resource-constrained environments <ref:2608.16769#pg1>.
Priya: What does this mean for the real world? Why is making it deployment oriented so important when we're talking about industrial control systems?
Nadia: Because these networks, like PLCs and SCADA systems, are increasingly targeted by distributed denial of service attacks that can physically disrupt processes or compromise safety <ref:2608.16769#pg1>.
Elias: The paper claims this framework addresses the severe constraints of OT environments by fusing a GRU and a decision tree whose rule set stays fully interpretable <ref:2608.16769#pg2>. This is important because it balances pattern recognition with transparency.
Priya: I'm interested in what they do to make sure this doesn't just work in theory, but actually performs well on the data sets they used for testing
CIC-DDoS2019, Edge-IIoTset, and CICIoT23: .
Nadia: They apply a unified preprocessing pipeline to these three benchmarks. This pipeline includes label mapping, numeric feature selection, robust scaling, and training set only class balancing <ref:2608.16769#pg1>.
Elias: And for those imbalanced datasets like CIC-DDoS2019 and CICIoT23, they use training only undersampling of the attack class to get an approximately thirty percent benign to attack balance in the training set <ref:2608.16769#pg1>.
Priya: That undersampling is a common trick, but I wonder if that artificially balances things out and hides some real detection issues? What are the results they claim for this balanced approach?
Nadia: They report that in both cases, CIC-DDoS2019 and CICIoT23, the hybrid model attains ninety-nine point zero four percent accuracy with a mean classification confidence of zero point nine seven MCC <ref:2608.16769#pg1>.
Elias: And on the CICIoT23 dataset, they see ninety-eight point six one percent accuracy with an MCC of zero point seven six <ref:2608.16769#pg1>. They also note that this hybrid approach reduces the false negative rate below both the pure neural and pure symbolic baselines in both test cases <ref:2608.16769#pg1>.
Priya: Reducing the false negative rate is critical in security because a missed attack means actual harm occurs, so how does this reduction compare to those other methods?
Nadia: The paper shows that the fusion provides a lower false negative rate at a controlled false positive cost <ref:2608.16769#pg2>. This suggests the hybrid design is specifically tuned for OT priorities where an attack being missed is worse than a false alarm.
Elias: So, to sum up what we've heard about "A Deployment-Oriented and Resource-Efficient Neuro-Symbolic Framework for Explainable DDoS Detection in Operational Technology Networks", it’s a hybrid
Conclusion: Nadia: So we've looked at how they fused that gated recurrent unit with the decision tree for these industrial networks, and now we're getting to the final wrap-up of this paper on "A Deployment-Oriented and Resource-Efficient Neuro-Symbolic Framework for Explainable DDoS Detection in Operational Technology Networks".
Elias: Yeah, so it really boils down to making this complex detection system something you can actually run on a small industrial gateway or PLC without needing a super powerful computer.
Priya: It’s about taking the high accuracy of neural networks and pairing it with the clear rules of symbolic systems, which is exactly what OT engineers need for trust.
Nadia: Exactly, and the authors did this by doing late fusion—meaning you let both parts make their guesses separately before they combine them—and they optimized those combining parameters to keep the false negative rate low while keeping false alarms managed.
Elias: I see how that trade-off is key in this sector, because a missed attack can cause real physical damage, so getting that error profile right is a major deal.
Priya: But what does this actually mean for someone who just runs a plant on the floor? Does it make their job easier or harder to trust the alert?
Nadia: It makes it easier because the symbolic part gives them rules they can read and understand, and that keeps them from getting overwhelmed by just raw probability scores.
Elias: The authors even put in some validation testing on actual hardware—a Siemens gateway and a Raspberry Pi—and it runs incredibly fast, under one millisecond for inference.
Priya: So the big picture here is that we can get high-quality, explainable security in places where traditional security tools just can't fit because of the size and processing power limitations.
Nadia: Right, it’s about proving that you don't need a super expensive machine to have a smart defense system for critical infrastructure.
Elias: And this sets up some interesting questions about how we build these neuro-symbolic systems across different distributed sites.
Keele University · University of Liverpool · University of Medea · The University of Electro-Communications, Tokyo, Japan
cs.CR
Submitted: 2026-08-17
Updated: 2026-11-17
Comments: 15 Version 2
License: http://creativecommons.org/licenses/by/4.0/
The gist: A deployment-oriented neuro-symbolic framework for explainable DDoS detection in operational technology networks presents a resource-efficient solution by fusing a compact GRU with an interpretable…
Key concepts
- Neuro-Symbolic Framework
- This approach blends neural networks, which are good at finding complex patterns in data, with symbolic reasoning, which uses human-readable rules. The system uses a GRU for pattern recognition and a decision tree for clear explanations of its decisions.
- GRU (Gated Recurrent Unit)
- A type of recurrent neural network used here to analyze time-series data, like network traffic flows. It is designed to capture temporal attack patterns, meaning it understands how the sequence of events over time contributes to identifying a DDoS attack.
- Late Score Fusion
- This is the method used to combine the outputs from the neural and symbolic parts. Instead of deeply integrating them during training, they generate separate scores, which are then combined later using a weighted average. This allows each component to operate independently while leveraging their strengths.
- Operational Technology (OT) Constraints
- This refers to industrial control systems like PLCs and SCADA used in factories. These environments have strict limitations on processing power, memory, and latency, which is why the model must be extremely fast and lightweight for deployment.
Terminology
Summary
A deployment-oriented neuro-symbolic framework for explainable DDoS detection in operational technology networks presents a resource-efficient solution by fusing a compact GRU with an interpretable decision tree to address the severe constraints of OT environments The hybrid attains 99. p. 0 Garcia-Luna-Aceves
Abstract—Operational technology (OT) environments, programmable logic controllers (PLCs), industrial control systems
(ICS), and supervisory control and data acquisition (SCADA)
systems, are increasingly targeted by distributed denial-of-service
(DDoS) attacks that can disrupt physical processes, damage
equipment, and compromise safety, while OT edge devices impose
severe computation, memory, and latency constraints. This paper presents a deployment-oriented neuro-symbolic framework
for DDoS detection in such resource-constrained environments,
fusing a compact gated recurrent unit (GRU) with a shallow
decision tree whose rule set remains fully interpretable. A
unified preprocessing pipeline (label mapping, numeric feature
selection, robust scaling, and training-set-only class balancing)
is applied to three benchmarks (CIC-DDoS2019, Edge-IIoTset,
and CICIoT23), and the fusion weight α and decision threshold
τ are jointly optimised on validation data. Because the held-out
test splits are attack-dominant, accuracy is reported alongside
the Matthews correlation coefficient (MCC) and per-class falsepositive and false-negative rates (FPR/FNR). This paper presents a deployment-oriented neuro-symbolic framework
for DDoS detection in such resource-constrained environments,
fusing a compact gated recurrent unit (GRU) with a shallow
decision tree whose rule set remains fully interpretable. A
unified preprocessing pipeline (label mapping, numeric feature
selection, robust scaling, and training-set-only class balancing)
is applied to three benchmarks (CIC-DDoS2019, Edge-IIoTset,
and CICIoT23), and the fusion weight α and decision threshold
τ are jointly optimised on validation data. A
unified preprocessing pipeline (label mapping, numeric feature
selection, robust scaling, and training-set-only class balancing)
is applied to three benchmarks (CIC-DDoS2019, Edge-IIoTset,
and CICIoT23), and the fusion weight α and decision threshold
τ are jointly optimised on validation data. Because the held-out
test splits are attack-dominant, accuracy is reported alongside
the Matthews correlation coefficient (MCC) and per-class falsepositive and false-negative rates (FPR/FNR). 61%
(MCC 0. 76) on CICIoT23, in both cases reducing the FNR
below the pure-neural and pure-symbolic baselines; on the
linearly separable Edge-IIoTset the tree alone reaches 100%,
validating the pipeline rather than the fusion. Model-only inference is submillisecond (0. 79 ms) on a workstation CPU, and hardwarein-the-loop validation of an INT8 deployment on a Siemens
SIMATIC IoT2050 industrial gateway and a Raspberry Pi 5
measures 2.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
System Architecture
The proposed neuro-symbolic DDoS detection system is designed for deployment on resource-constrained OT edge devices such as PLCs, industrial gateways, and SCADA frontends It consists of four main stages:
-
Data preprocessing: raw network flows are cleaned, numeric features are selected, missing values are imputed, and the data are normalised with a RobustScaler
-
Sliding-window formation: consecutive feature vectors are grouped into windows of length L = 16; the full window feeds the neural component and only the last vector feeds the symbolic component
-
Parallel neural and symbolic inference: a two-layer GRU processes the whole window to capture temporal attack patterns and outputs pneural, while a shallow decision tree processes the last feature vector and outputs psymbolic from human-readable rules
-
Fusion and decision: the probabilities are combined via a convex weight α as phybrid = αpneural+ (1−α)psymbolic, and a threshold τ produces the final binary classification
How the training signal is built
The system employs a unified preprocessing pipeline formalized in Algorithm 1, which includes keyword-based binary label mapping, retention of numeric features only, and stratified 70%/15%/15% train/validation/test split For imbalanced datasets like CIC-DDoS2019 and CICIoT23, the training-only undersampling of the attack class to an approximately 30% benign-to-attack balance is applied to Xtrain only The parameters α and τ are jointly optimised on the validation set to maximise the F1-score by an exhaustive grid search over α ∈ [0, 1] and τ ∈ (0, 1)
Where the gains come from
The principal gain of the fusion is a lower FNR at a controlled false-positive cost, which is the operative trade-off in OT where a missed attack is more damaging than a false alarm The hybrid provides the best balance on base-rate-aware metrics (MCC and FNR) on the two non-trivial benchmarks, consistent with the OT priority that a missed attack is more damaging than a false alarm The symbolic branch supplies the interpretability OT analysts require while the neural branch handles complex, imbalanced patterns
Inference Performance and Validation
The total inference latency is essentially the maximum of the two individual latencies plus the weighted combination, and all variants achieve sub-millisecond inference Hardwarein-the-loop validation on a Siemens SIMATIC IoT2050 and a Raspberry Pi 5 confirms millisecond-scale INT8 inference, at most 31 MB resident memory, and millijoule-scale energy per inference, well within the 50 ms OT control-loop budget The hybrid model achieves sub-millisecond inference latency below 0.8 ms per sample on a standard CPU
Performance Comparison and Limitations
The hybrid provides competitive detection performance (98–100% accuracy and 99–100% F1) while delivering sub-millisecond inference latency, a deployment-oriented metric that is unavailable or unreported in most existing neuro-symbolic IDS studies The higher FPR on CICIoT23 (8.5%) suggests that further calibration or a different balancing strategy, such as minorityclass oversampling, is needed to reduce false alarms without increasing missed attacks The coupling is loose (late score fusion), not tightly coupled differentiable logic; the symbolic component adds interpretable rules and a complementary error profile but does not constrain the neural optimisation, so symbolicreasoning claims are limited to decision-tree rule extraction The higher FPR on CICIoT23 (8.5%) suggests that further calibration or a different balancing strategy, such as minorityclass oversampling, is needed to reduce false alarms without increasing missed attacks
Conclusion and Outlook
This work shows that a minimal, auditable latefusion neuro-symbolic design can approach the accuracy of heavier architectures while meeting OT resource and latency constraints Future work will extend the hardware-in-loop validation of Section VII to physically commissioned PLC endpoints and to additional SoC, runtime, and quantisation combinations Broader deployment scenarios may benefit from recent advances in multi-agent and federated neuro-symbolic architectures enabling collaborative threat-intelligence sharing across distributed OT sites while preserving data privacy
REFERENCES
[1] A. Mustapha, R. Khatoun, S. Zeadally, F. Chbib, A. Fadlallah, W. Fahs, and A. El Attar, Detecting DDoS attacks using adversarial neural network,
Comput. Secur., vol. 127, Art no 103117, 2023
[4] S. Dey, A. Paul, A. Mukherjee, D. Basu, and U Ghosh, DeNSAINet: CNN + symbolic rules,
in Proc IEEE 6th India Council Int Subsections Conf (INDISCON), Rourkela, India Apr 2025
[8] A. Bizzarri, B Jalaian, F Riguzzi, and N D Bastian, A neuro-symbolic artificial intelligence network intrusion detection system,
in Proc 33rd Int Conf Comput Commun Netw (ICCCN), Kailua-Kona, HI USA Jul 2024
[9] D Onchis, C Istin, and E Hogea, A neuro-symbolic classifier with optimized satisfiability for monitoring security alerts in network traffic,
Appl Sci., vol 12 Art no 11502, 2022
[13] A Nerella, P Badri, S T Reddy Kandula, V Reddy Surasani, P K Muthukamatchi, and A Jain, Neurosymbolic AI for IoT security: A knowledge-guided framework for real-time IoT anomaly detection and response,
in Proc 17th Int Conf Contemp Comput (IC3), Noida, India Aug 2025
[15] H T T Tran, J Sander, A Cohen, B Jalaian, and N D Bastian, Neurosymbolic artificial intelligence for robust network intrusion detection: From scratch to transfer learning,
2025
[18] S Nalluri, M M Malyala, H Kandagiri, and K K Kandagiri, NSCTI: A hybrid neuro-symbolic framework for AI-driven predictive cyber threat intelligence,
in Proc 4th Int Conf Comput Model Simul Optim (ICCMSO), Singapore Feb 2025
[19] F Ares-Robledo, H Rifà-Pous, and R Clarisó, Hyperparameter optimization in neuro-symbolic unsupervised graph learning,
in Availability Reliability and Security (ARES 2025), Lecture Notes in Computer Science, vol 15997, B Coppens, B Volckaert, V Naessens, and B De Sutter Eds.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Resource-constrained deployment feasibility: The system can achieve millisecond-scale inference on industrial hardware, confirming suitability for deployment on
resource-constrained OT edge gateways, PLCs, and SCADA front-end processors.
-
Explainable detection mechanism: The hybrid model provides human-readable rules via the symbolic branch (
the tree yields human-readable rules (e.g., if flow duration exceeds a threshold and packet length is below a threshold, then DDoS)
), buildingoperator trust and facilitating forensic analysis.
-
Adaptive operating regime: The joint optimization of the fusion weight alpha and decision threshold tau allows the system to be tuned per OT environment, meaning
the symbolic component alone may suffice in balanced settings, while the neural component takes precedence in highly imbalanced or complex ones.
-
Temporal pattern capture: The GRU component processes a sliding window of 16 vectors (
The GRU computations are given by (2)–(5)
) tocapture temporal attack patterns,
which is essential for detecting attacks likeSlowITe, a low-rate attack exploiting the MQTT keep-alive mechanism.
-
Performance on imbalanced data: The framework utilizes
training-only undersampling of the imbalanced datasets
and achieves strong base-rate metrics like MCC 0.97 on CIC-DDoS2019, demonstrating its ability to perform well even when attack classes are under-sampled in training sets. -
Operational trade-off control: The system manages the critical OT trade-off by achieving
a lower FNR at a controlled false positive cost,
specifically addressing the OT priority thata missed attack is more damaging than a false alarm.
Abstract
This paper presents a hybrid neural-symbolic detector of distributed denial-of-service (DDoS) attacks on the availability of operational technology (OT) networks, designed for their resource-constrained gateways, and an evaluation matched to deployment. A compact gated recurrent unit (GRU) is fused at score level with a shallow, readable decision tree, and weight and threshold are selected on validation data under an explicit false-positive budget, whose attainment in finite samples is analysed. The evaluation on CIC-DDoS2019, Edge-IIoTset and CICIoT23 uses contiguous or capture-held-out partitions, dependence-aware intervals, and precision and alert load at operational prevalence. Over ten seeds at a 1% budget, the hybrid's false-negative rate on CIC-DDoS2019 is 0.08%, against 0.26% for the GRU and 0.17% for the tree. With a day of mostly unseen attack families held out, the tree alone carries the decision. Audits bound these claims: Edge-IIoTset is saturated by duplicates, no benchmark as used provides interleaved traffic in capture order, and an unconstrained tree is more accurate on two of them, so the case for the hybrid rests on short, stable rules and a measured footprint. On a Siemens SIMATIC IoT2050 gateway, tested with Modbus/TCP and MQTT floods, the INT8 model needs 2.9 ms per window, 31 MB of memory and 3.5 mJ per inference.
Sources
- A Synergistic Approach In Network Intrusion Detection By Neurosymbolic AI
- Neurosymbolic Artificial Intelligence for Robust Network Intrusion Detection: From Scratch to Transfer Learning
- A Survey for Deep Reinforcement Learning Based Network Intrusion Detection
Related papers
- From Strategy Search to Policy Learning: MCTS-Guided Adversarial Co-Evolution for LLM Safety Alignment
- Privacy Leakage from Recurrent State in LLM Inference
- Incrementally Verifiable Oracle Aided Computation
- Privilege and confidentiality in generative AI workflows
- BadMoE: Backdooring Mixture-of-Experts LLMs via Routing-Aware Triggers and Task-Underutilized Experts
- CoVault: Secure, Scalable Analytics of Personal Data