Explainability-Aware Evaluation of Transfer Learning Models for IoT DDoS Detection Under Resource Constraints

Nelly Elsayed

cs.CRFeb 25, 2026

Explainability-Aware Evaluation of Transfer Learning Models for IoT DDoS Detection Under Resource Constraints

Authors: Nelly Elsayed

Organizations: School of Information Technology, University of Cincinnati, Ohio, United States

Abstract

Distributed denial-of-service (DDoS) attacks threaten the availability of Internet of Things (IoT) infrastructures, particularly under resource-constrained deployment conditions. Although transfer learning models have shown promising detection accuracy, their reliability, computational feasibility, and interpretability in operational environments remain insufficiently explored. This study presents an explainability-aware empirical evaluation of seven pre-trained convolutional neural network architectures for multi-class IoT DDoS detection using the CICDDoS2019 dataset and an image-based traffic representation. The analysis integrates performance metrics, reliability-oriented statistics (MCC, Youden Index, confidence intervals), latency and training cost assessment, and interpretability evaluation using Grad-CAM and SHAP. Results indicate that DenseNet and MobileNet-based architectures achieve strong detection performance while demonstrating superior reliability and compact, class-consistent attribution patterns. DenseNet169 offers the strongest reliability and interpretability alignment, whereas MobileNetV3 provides an effective latency-accuracy trade-off for fog-level deployment. The findings emphasize the importance of combining performance, reliability, and explainability criteria when selecting deep learning models for IoT DDoS detection.

Figures & tables

Figure 1: The cybersecurity CIA triad that represents the three pillars of information security: confidentiality, integrity, and availability [ 25 ] .

Figure 2: The IoT three-layered architecture [ 76 ] .

Figure 3: The proposed research methodology for investigating pre-trained models for DDoS attack detection.

Figure 4: Network traffic data transformation into image representation for pre-trained models.

Model	Model Description	No. Param.	Type	Handle Input Size
MobileNet	Lightweight CNN for mobile devices	4M	Lightweight	Yes
MobileNetV2	Improved MobileNet with depthwise separable convolutions	3.4M	Lightweight	Yes
MobileNetV3	Optimized MobileNet version for efficiency	5.4M	Lightweight	Yes
DenseNet121	Densely connected CNN with fewer parameters	8M	CNN-Based	Yes
DenseNet169	Deeper DenseNet variant	14M	CNN-Based	No
EfficientNet-B0	Optimized for accuracy–efficiency balance	5.3M	Advanced Architecture	No

Table 1: Comparison of selected pre-trained models based on selection criteria.

Model	No. Layers	20% Unfreeze	Trainable Param.	Non-trainable Param.
MobileNet	89	17	1,924,404	1,371,840
MobileNetV2	157	33	1,609,164	731,584
MobileNetV3	279	58	2,907,564	1,401,632
DenseNet121	430	88	1,785,548	5,318,336
DenseNet169	598	122	3,780,428	8,969,792
EfficientNet-B0	240	50	2,388,060	1,744,275

Table 2: The fine-tune elements comparison between the selected pre-trained models for the empirical evaluation.

Model	Train Time (m)	Avg. test time/sample (s)
MobileNet	61.745	0.000530
MobileNetV2	109.633	0.000864
MobileNetV3	140.915	0.001093
DenseNet121	248.882	0.002191
DenseNet169	345.697	0.003050
EfficientNet-B0	183.157	0.001427

Table 3: A comparison between the pretrained models training and testing time.

Model	Accuracy	Precision	Recall	F1-Score	AUC (ROC)	SE	Kappa	HL
MobileNet	98.476%	0.90895	0.9008	0.90434	0.94621	0.00163	0.89711	0.09144
MobileNetV2	98.43%	0.9058	0.9058	0.89591	0.94364	0.00166	0.89399	0.0942
MobileNetV3	98.059%	0.89205	0.87209	0.88353	0.93506	0.00182	0.86870	0.11647
DenseNet121	98.25%	0.89142	0.88689	0.88814	0.93863	0.00174	0.88187	0.10498
DenseNet169	98.298%	0.89788	0.89007	0.89788	0.94036	0.00172	0.88509	0.10212
EfficientNet-B0	93.156%	0.58938	0.59623	0.58938	0.77878	0.00279	0.53711	0.41062

Table 4: A comparison between the pretrained models on DDoS attack detection.

Panel of Figure 5 — Figure 5: The train vesus validation for the pre-trained models for detecting the DDoS attacks.

Panel of Figure 6 — Figure 6: The ROC diagrams for the pre-trained models for detecting the DDoS attacks.

Model	95% CI	MCC	Balanced Acc	Log Loss	Youden Index
MobileNet	(0.90536,0.91177)	0.89718	0.9008	3.28369	0.89242
MobileNetV2	(0.90255, 0.90905)	0.89425	0.94364	3.28699	0.88728
MobileNetV3	(0.87997,0.88710)	0.87077	0.93063	3.32906	0.86127
DenseNet121	(0.89161, 0.89843)	0.88212	0.93863	3.28696	0.87726
DenseNet169	(0.89452, 0.90125)	0.88523	0.94036	3.28484	0.88071
EfficientNet-B0	(0.58391, 0.59485)	0.55983	0.77878	2.89804	0.55756

Table 5: A statistical analysis comparison between the pretrained models on DDoS attack detection.

Panel of Figure 7 — Figure 7: Grad-CAM heatmaps for the pre-trained models for detecting the DDoS attacks.

Model	Spatial Compactness	Class Consistency	Background Suppression	Interpretability Level
DenseNet169	Very High	Very High	Strong	Excellent
DenseNet121	High	High	Strong	Very Good
MobileNetV3	Moderate to High	Moderate to High	Moderate	Good
MobileNet	Moderate	Moderate	Moderate	Fair–Good
MobileNetV2	Moderate	Moderate	Moderate to Low	Moderate
EfficientNet-B0	Low to Moderate	Moderate	Weak	Moderate

Table 6: Grad-CAM Spatial Localization Comparison

Panel of Figure 8 — Figure 8: The SHAP results for the pre-trained models for detecting the DDoS attacks.

Model	SHAP Magnitude	Attribution Concentration	Negative Suppression	Attribution Strength
DenseNet169	$\pm$ 0.03	Very Compact	Strong	Very Strong
DenseNet121	$\pm$ 0.03	Compact	Strong	Strong
MobileNetV3	$\pm$ 0.02–0.025	Moderate	Moderate	Moderate–Strong
MobileNet	$\pm$ 0.02	Moderate	Moderate	Moderate
EfficientNet-B0	$\pm$ 0.02	Diffused	Moderate–Weak	Moderate
MobileNetV2	$\pm$ 0.01	Less Compact	Moderate–Weak	Weak–Moderate

Table 7: SHAP Attribution Comparison Across Architectures

Model	Rank
DenseNet169	1
DenseNet121	2
MobileNetV3	3
MobileNet	4
EfficientNet-B0	5
MobileNetV2	6

Table 8: Overall Interpretability Ranking Based on Grad-CAM and SHAP Analysis

Panel of Figure 9 — Figure 9: Confusion matrices for all evaluated pre-trained architectures on multi-class DDoS detection.

IoT Intrusion Detection

IoT intrusion detection identifies malicious network activity targeting connected devices and industrial systems using traffic and device data.

See all 51 papers

Beyond Detection Accuracy: Measuring Explanation Cost, Stability, and Utility for Resource-Aware IoT Intrusion Detection
Aug 11, 2026Abdurrahman TolaySHAP Feature Attribution IoT Intrusion Detection
Independent Researcher, Istanbul, Türkiye
Cross-Domain Generalization Failure in Lightweight Intrusion Detection Models for IIoT Networks
Jul 1, 2026MD Azizul Hakim, Md Shihab Uddin, Talha Ibne AnisIoT Intrusion Detection Network Intrusion Detection
Department of Computer Science and Technology, Bangladesh Sweden Polytechnic Institute Kaptai, Rangamati, Chittagong 4530, Bangladesh · Department of Electrical Engineering, Bangladesh Sweden Polytechnic Institute Kaptai, Rangamati, Chittagong 4530, Bangladesh
Explainable Threat Attribution for IoT Networks Using Conditional SHAP and Flow Behavior Modelling
Mar 24, 2026Samuel Ozechi, Jennifer OkonkwoabutuExplainable Intrusion Detection IoT Intrusion Detection
University of East London, United Kingdom · ECE Ecole D’Ingenieur, France

Aug 11, 2026J/K move · Enter open · S save

Beyond Detection Accuracy: Measuring Explanation Cost, Stability, and Utility for Resource-Aware IoT Intrusion Detection

Abdurrahman Tolay

Independent Researcher, Istanbul, Türkiye

SHAP Feature Attribution IoT Intrusion Detection Cost-Aware Inference Explanation Stability Explainability Evaluation Explainable Intrusion Detection

Open paper AskPDF arXiv

Machine-learning intrusion-detection studies commonly emphasize predictive accuracy while treating explanation generation as a computationally free post-processing step. This study jointly evaluates predictive effectiveness, explanation cost, local explanation stability, and selective explanation for binary Internet of Things (IoT) intrusion detection. A leakage-safe CICIoT2023 corpus was constructed using exact 39-feature hashes, non-finite-value handling, exact-feature deduplication, conservative label-collision removal, and deterministic hash-level partitioning. Logistic Regression, Decision Tree, Random Forest, and XGBoost were evaluated on natural and balanced test distributions. TreeSHAP cost was measured, stability was assessed under prediction-preserving perturbations, and validation-calibrated policies were used to allocate explanation workload. XGBoost provided the strongest overall predictive profile, while Random Forest produced the lowest false-positive rate. At 5,000 samples, TreeSHAP required 700.759 s for Random Forest and 1.471 s for XGBoost. Random Forest showed the strongest overall base-level explanation stability; XGBoost retained high rank and directional consistency but showed greater top-feature turnover and attribution-magnitude drift. On the balanced test, about 90% false-negative explanation coverage permitted 28-32% compute savings, while about 95% coverage permitted 15-23% savings. Savings were much smaller under the attack-heavy natural prevalence. These results show that operationally useful explainable IoT intrusion detection depends on predictive quality, explanation cost, local stability, workload prevalence, and selective invocation rather than detection accuracy alone.

Abstract

IoT Intrusion Detection

Explore similar work

Beyond Detection Accuracy: Measuring Explanation Cost, Stability, and Utility for Resource-Aware IoT Intrusion Detection

Cross-Domain Generalization Failure in Lightweight Intrusion Detection Models for IIoT Networks

Explainable Threat Attribution for IoT Networks Using Conditional SHAP and Flow Behavior Modelling