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

arXiv:2602.22488 · cs.CR, cs.AI · Submitted 2026-02-25 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: I'm Nadia.

Elias: And I'm Elias, here with Priya.

Nadia: We're looking at the paper: "Explainability-Aware Evaluation of Transfer Learning Models for IoT DDoS Detection Under Resource Constraints".

Elias: Here's the gist of the paper. Distributed denial-of-service (DDoS) attacks threaten the availability of Internet of Things (IoT) infrastructures, particularly under resource-constrained deployment conditions.

Nadia: 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 imagebased traffic representation.

Elias: 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.

Nadia: Results indicate that DenseNet and MobileNet-based architectures achieve strong detection performance while demonstrating superior reliability and compact, class-consistent attribution patterns.

Paper summary: Nadia: The paper sets up a threat model focusing on DDoS attacks targeting IoT networks where defenders monitor network flow-level features instead of doing deep packet inspection >

Elias: The primary security goal they set is to preserve service availability under these DDoS conditions, aiming for timely and accurate identification across different attack types >

Nadia: They used the CICDDoS2019 dataset and turned tabular traffic data into sixty by sixty by three image representations for the models to classify > <ref:2602.22488#pg3>

Priya: So they're taking this complex traffic data and turning it into pictures so a standard computer vision model can look at it, which is a smart way to approach pattern recognition >

Elias: The thesis of the study is that they are evaluating explainability-aware architectures because traditional models lack transparency when deployed in these constrained IoT settings >

Conclusion: Nadia: So, looking at the title, Explainability-Aware Evaluation of Transfer Learning Models for IoT DDoS Detection Under Resource Constraints, it really boils down to checking if these models are not just accurate but also trustworthy when they run on limited hardware >

Elias: The authors found that DenseNet and MobileNet-based architectures showed strong detection performance alongside superior reliability and compact attribution patterns >

Priya: That means for a deployment scenario where you can't afford a huge model, these two types are the most promising starting points because they balance good results with being reasonably efficient >

Nadia: The implication here is that for IoT defenders operating at the edge, choosing between DenseNet169 and MobileNetV3 depends on what you need: accountability or low latency >

Elias: If you need decision accountability where you can trust *why* the model flagged something, DenseNet169 seems strongest because of its compact localization and strong SHAP attribution magnitudes >

Priya: But if the main concern is keeping things fast on a very small device, MobileNetV3 is recommended as the most lightweight option that still maintains good interpretability coherence >

School of Information Technology, University of Cincinnati

cs.CR, cs.AI

Submitted: 2026-02-25

Updated: 2026-10-30

Comments: Updated version with several corrections, contains more clarification, and corrected typos in at the text and references

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: Distributed denial-of-service (DDoS) attacks threaten the availability of Internet of Things (IoT) infrastructures, particularly under resource-constrained deployment conditions.

Terminology

Summary

Distributed denial-of-service (DDoS) attacks threaten the availability of Internet of Things (IoT) infrastructures, particularly under resource-constrained deployment conditions. p. 0 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 imagebased traffic representation. p. 0 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. p. 0 Results indicate that DenseNet and MobileNet-based architectures achieve strong detection performance while demonstrating superior reliability and compact, class-consistent attribution patterns. p. 0

Threat Context and Objectives

The paper establishes a threat model focusing on distributed denial-of-service (DDoS) attacks targeting IoT networks. The defender operates at the fog or gateway level to monitor network flow-level features without deep packet inspection. p. 2 The primary security goal is to preserve the availability of IoT-driven systems under DDoS conditions. p. 2 Specifically, the detection mechanism aims to provide timely and accurate identification of attack traffic while maintaining stability across heterogeneous attack categories. p. 2

Data Transformation and Model Selection

The research utilizes the CICDDoS2019 benchmark dataset for evaluation. Structured tabular traffic data is transformed into image-based representations by extracting consecutive chunks of 180 traffic records and reshaping them into RGB images of dimension 60 × 60 × 3. p. 4 Pre-trained models were selected based on criteria such as being designed for image classification within transfer learning frameworks and having a total number of parameters below 15 million to align with resource constraints. p. 4 The fine-tuning strategy involves partially unfreezing the top layers of each model while keeping the remaining layers frozen to support domain adaptation. p. 5

Performance and Reliability Metrics

Model performance is measured using standard classification metrics like accuracy, precision, recall, F1-score, and AUC. Beyond accuracy, reliability is assessed using Cohen’s kappa and Hamming loss. p. 6 Confidence-oriented statistical measures such as 95% confidence intervals, Matthews correlation coefficient (MCC), balanced accuracy, log loss, and Youden’s index are computed to characterize stability and robustness under multi-class conditions. p. 6 DenseNet169 and MobileNet achieve strong MCC values near 0. p. 0885–0.897 and high Youden Index values near 0. p. 988–0.89, supporting their suitability for operational deployment. p. 0

Interpretability Analysis and Findings

Explainability is assessed using Grad-CAM for spatial localization and SHAP for quantitative feature attribution. DenseNet architectures demonstrate the highest spatial compactness and class consistency, with strong background suppression. p. 9 DenseNet169 consistently achieves the highest spatial compactness, the strongest SHAP attribution magnitude, and the most effective background suppression. p. 0 MobileNetV3 represents the strongest lightweight architecture in terms of interpretability, while EfficientNet-B1 exhibits the lowest degree of discriminative localization and feature decisiveness. p. 9

Deployment Recommendations

The findings suggest a pragmatic deployment strategy based on operational needs. MobileNetV3 is recommended for low-latency edge detection due to its balance between efficiency and explainability. p. 5 DenseNet169 is the strongest choice for reliability-critical pipelines where decision accountability and analyst trust are central requirements. p. 6 EfficientNet architectures are less favorable due to reduced stability, weaker reliability, and lower interpretability clarity. p. 9

Conclusion

The results demonstrate that MobileNet- and DenseNet-based architectures achieve strong multi-class detection performance (approximately 98% accuracy), with high agreement and reliability indicators, including Cohen’s Kappa, MCC, and Youden Index. DenseNet169 emerges as the most reliable and explainable architecture, exhibiting compact, class-consistent Grad-CAM localization and the strongest SHAP attribution magnitudes. p. 0 MobileNetV3 represents the strongest lightweight alternative, providing high detection capability with improved interpretability coherence compared to MobileNetV2. p. 5 In contrast, EfficientNet-B0 and especially EfficientNet-B1 exhibit reduced stability, weaker reliability statistics, and more diffuse interpretability patterns under the evaluated configuration. p. 9

Future Work

Future work will extend this study by incorporating deeper, confusion-matrix-driven class-specific error analysis, evaluating robustness under distribution shifts (e.g., unseen attack variants), and exploring edge-aware compression and quantization strategies to reduce inference cost further while preserving interpretability and reliability guarantees. p. 0

Credit Authorship Contribution Statement

Nelly Elsayed (sole author): Conceptualization, Methodology, Data Curation, Formal Analysis, Investigation, Visualization, Writing, and Original Draft Preparation. The authors used generative AI including ChatGPT, Grammarly, and Gemini tools solely for language refinement to ensure that the texts are free of errors in grammar, spelling, and punctuation. p. 9

Declaration of Competing Interest

The authors declare that they have no competing financial or personal interests that could have influenced the work reported in this paper. The dataset CIC-DDoS2019 Dataset is publicly available at the following links. p. 3 The image format of this dataset is publicly available but requires membership to access using the IEEE DataPort. p. 3 p. 3 p. 3 The paper was prepared using LaTeX formatting assistance. p. 1 p. 1 p. 1

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-forprofit sectors. The paper was prepared using Gemini tools solely for language refinement to ensure that the texts are free of errors in grammar, spelling, and punctuation. p. 3 The authors take full responsibility for the content and integrity of the final manuscript. p. 5

References

The references listed support the context and related work discussed in the paper. The authors cite various works regarding IoT, DDoS threats, and machine learning approaches for detection. p. 1 The dataset CIC-DDoS2019 is publicly available at the following links. p. 3 p. 3 The image format of this dataset is publicly available but requires membership to access using the IEEE DataPort. The paper was prepared using LaTeX formatting assistance.

The authors cite various works regarding IoT, DDoS threats, and machine learning approaches for detection. The dataset CIC-DDoS2019 is publicly available at the following links. The image format of this dataset is publicly available but requires membership to access using the IEEE DataPort. The paper was prepared using LaTeX formatting assistance.

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

  1. Improve model selection by prioritizing DenseNet169 for reliability-critical systems, as it demonstrates the highest spatial compactness and class consistency in Grad-CAM analysis, ensuring more defensible and transparent detection behavior.

  2. Implement a dual-path deployment strategy where MobileNetV3 is used for low latency fog-level filtering due to its lowest inference time per sample (0.000530 s), while DenseNet169 is reserved for high-stakes scenarios requiring analyst trust and decision transparency.

  3. Enhance the detection system's decision confidence by integrating a reliability score derived from the Youden Index or Matthews Correlation Coefficient, as these metrics provide evidence that models with high MCC and Youden Index values provide more trustworthy multi-class decision behavior.

  4. Develop an automated model monitoring system that flags performance degradation based on the stability observed in training dynamics, specifically identifying architectures like EfficientNet-B0 and B1 which exhibit noticeable validation oscillations and larger train–validation gaps.

  5. Create a real-time explainability module using a combination of Grad-CAM for spatial localization and SHAP for feature attribution to generate clear positive–negative separation in attack signatures, allowing security analysts to understand the decisive feature reliance of the detection system.

Sources

Related papers