Adaptive multi-model knowledge distillation with dynamic teacher selection for efficient lightweight network intrusion detection
Abstract & Details
Research Area
Computer sciences, Artificial intelligence
Keywords
Network Intrusion Detection
Knowledge Distillation
Multi-Teacher Learning
Adaptive Learning
IA
Abstract
Network intrusion detection is a major challenge for digital infrastructure security, particularly when computational and memory resources are limited. Artificial intelligence models offer promising capabilities for improving detection performance; however, their complexity can make deployment difficult in resource-constrained systems. This article proposes ADMS-KD, an adaptive and dynamic multi-model system based on knowledge distillation for lightweight network intrusion detection. The method aims to leverage complementary knowledge from several teacher models, including CNN, LSTM, MLP, and Transformer, and transfer their decision-making capabilities to a less complex student model. The proposed approach uses flow features from a hypothetical dataset, followed by preprocessing and normalization. The different teacher models then generate output probability distributions. These predictions are combined through an adaptive and dynamic mechanism to construct a shared knowledge representation, which is subsequently transferred to the student model through distillation. The results assess the student model’s computational footprint and inference cost. They indicate that the proposed system can maintain high classification performance while reducing the computational costs associated with the final model. Comparison with conventional distillation also highlights the benefits of jointly exploiting knowledge from multiple teacher models. Furthermore, the results suggest that the student model retains a strong ability to distinguish normal traffic from the intrusion categories considered. These findings underscore the potential of ADMS-KD for designing lightweight network intrusion detection systems that combine effective detection, low computational cost, and strong generalization capabilities. The adaptive selection of teacher contributions may help balance predictive reliability and computational efficiency, supporting deployment in resource-constrained digital infrastructures across varied network conditions.
License
This work is licensed under a Creative
Commons
Attribution-ShareAlike 4.0 International License.
Author Information
| # | Name | Institute / Affiliation |
|---|---|---|
| 1 | RAMAHEFY Tiana Razefania | University of Vakinankaratra |
| 2 | RANAIVOSOA Rija | University of Vakinankaratra |
| 3 | RABEHERIMANANA Toussaint | University of Antananarivo |
| 4 | RASOLOMANANA Eddy Harilala | University of Antananarivo |
How to Cite
Use the following formats to cite this article in your research.
APA Style
Razefania, RAMAHEFY Tiana, Rija, RANAIVOSOA, Toussaint, RABEHERIMANANA, & Harilala, RASOLOMANANA Eddy (2026). Adaptive multi-model knowledge distillation with dynamic teacher selection for efficient lightweight network intrusion detection. International Journal of Advance Research and Innovative Ideas In Education, 12(5), 752-766.
MLA Style
Razefania, RAMAHEFY Tiana, et al. "Adaptive multi-model knowledge distillation with dynamic teacher selection for efficient lightweight network intrusion detection." International Journal of Advance Research and Innovative Ideas In Education, vol. 12, no. 5, 2026, pp. 752-766.
IEEE Style
RAMAHEFY Tiana Razefania, RANAIVOSOA Rija, RABEHERIMANANA Toussaint, and RASOLOMANANA Eddy Harilala, "Adaptive multi-model knowledge distillation with dynamic teacher selection for efficient lightweight network intrusion detection," International Journal of Advance Research and Innovative Ideas In Education, vol. 12, no. 5, pp. 752-766, 2026.
Vancouver Style
Razefania RAMAHEFY Tiana, Rija RANAIVOSOA, Toussaint RABEHERIMANANA, Harilala RASOLOMANANA Eddy. Adaptive multi-model knowledge distillation with dynamic teacher selection for efficient lightweight network intrusion detection. International Journal of Advance Research and Innovative Ideas In Education. 2026;12(5):752-766.
Harvard Style
Razefania, RAMAHEFY Tiana, Rija, RANAIVOSOA, Toussaint, RABEHERIMANANA, & Harilala, RASOLOMANANA Eddy (2026) 'Adaptive multi-model knowledge distillation with dynamic teacher selection for efficient lightweight network intrusion detection', International Journal of Advance Research and Innovative Ideas In Education, 12(5), pp. 752-766.
Chicago Style
Razefania, RAMAHEFY Tiana, et al. "Adaptive multi-model knowledge distillation with dynamic teacher selection for efficient lightweight network intrusion detection." International Journal of Advance Research and Innovative Ideas In Education 12, no. 5 (2026): 752-766.
Turabian Style
Razefania, RAMAHEFY Tiana, et al. "Adaptive multi-model knowledge distillation with dynamic teacher selection for efficient lightweight network intrusion detection." International Journal of Advance Research and Innovative Ideas In Education 12, no. 5 (2026): 752-766.
Related Research
Precision Agriculture Using Computer Vision: A Review of In-Field Disease Detection, Foliar Severity Estimation
Download PDF
Development of a Custom Handwritten Kannada Dataset and CNN-BiLSTM-CTC Based OCR System
Download PDF
CodeCoach AI: A Review of Automated Programming Assessment, Explainable Feedback, and an AI-Assisted Coding Practice Framework
Download PDF
BudgetBee - A Review of AI Enabled Personal Finance Management and Smart Budgeting Application
Download PDF
Krishi Sathi and AI-Powered Farming Assistants: A Systematic Review of Conversational and Decision-Support Intelligence for Smallholder Agriculture
Download PDF
EVOLUTION OF SECOND-HAND BOOK SELLING: A SYNTHESIS OF TRADITIONAL MARKETS, DIGITAL PLATFORMS, AND INTELLIGENT TECHNOLOGIES
Download PDF