Lightweight Hybrid
DNN-GNN NIDS

Adaptive late fusion for real-time network intrusion detection on edge and IoT-class infrastructure.

ICNWC 2026 · Apr 6–8 — peer-reviewed — network intrusion detection — network security and deep learning

document

abstract

With the evolution of network architectures towards decentralized IoT and edge computing, the computational complexity involved in deep learning-based Network Intrusion Detection Systems (NIDS) has raised significant concerns about scalability. Graph Neural Networks (GNNs) are capable of representing relational graph structures in network traffic; yet, their reliance on the construction of graph structures makes them less practical in resource-limited settings. This paper proposes a new hybrid network intrusion detection system (NIDS) design that combines a Deep Neural Network (DNN) with a relational multi-layer perceptron (MLP) component, connected through an adaptive late fusion strategy that uses learnable fusion weights.Based on the latest findings that MLPs are capable of learning relational patterns from tabular data, the proposed methodology overcomes the need for the explicit construction of graphs while maintaining complementary representations of features. The empirical assessment on the CIC-IDS-2017 dataset shows an accuracy of 99.64% with around 80,000 trainable parameters. These findings indicate a favorable trade-off between accuracy and complexity, thus validating the appropriateness of the proposed architecture for real-time analysis in resource-constrained networks.een accuracy and complexity, thus validating the appropriateness of the proposed architecture for real-time analysis in resource-constrained networks.

key results

            99.64%
            Classification accuracy on the CIC-IDS-2017 benchmark
          

            80K
            Approximate trainable parameter budget for the hybrid model
          

            1,250x
            Fewer parameters than the Transformer baseline
          

            16K+
            Network flows per second on consumer hardware
          

authors

Aryan Mondal Dept. of Networking and Communications · SRM Institute of Science and Technology · Kattankulathur, India
Abhradeep Basak Dept. of Networking and Communications · SRM Institute of Science and Technology · Kattankulathur, India
Aniket Ghosh Khoury College of Computer Sciences · Northeastern University · Boston, MA, USA
M Lakshmi Dept. of Networking and Communications · SRM Institute of Science and Technology · Kattankulathur, India

keywords · index terms

intrusion detection systems graph neural networks adaptive fusion edge computing internet of things lightweight deep learning

venue · publication

conference ICNWC 2026

dates 6–8 April 2026

conference record IEEE # 68145

isbn 979-8-3195-4298-4

doi 10.1109/ICNWC68145.2026.11518522

document id IEEE Xplore · 11518522

pages 1 – 7

full title Lightweight Hybrid DNN-GNN Architecture for Network Intrusion Detection with Adaptive Late Fusion

read in full

Full text, architecture details, experimental setup, tables, and the complete reference list are hosted on IEEE Xplore. This page is a landing summary — the canonical version of the paper lives there.

open on IEEE Xplore ↗ resolve via DOI ↗ ← back to publications

Lightweight Hybrid DNN-GNN NIDS