Direct Current Feature Injection and Adaptive Dynamic Thresholding for Intrusion Detection using Leaky Integrate-and-Fire Neurons

Abstract Intrusion detection systems deployed at the network edge should detect attacks accurately under severe class imbalance while operating within tight energy budgets. Spiking neural networks (SNNs) offer event-driven, low-power, and sparse computation. However, their accuracy is limited by three factors: lossy spike encoding of real-valued traffic features, uncontrolled firing under fixed thresholds, and poorly separable input representations. This paper proposes an SNN-based intrusion detection model that resolves the mentioned limitations. First, a class-aware discriminative autoencoder, trained with a staged schedule of reconstruction pre-training, decoder freezing, and class-weighted fine-tuning, maps PCA-reduced traffic features to compact embeddings that preserve minority-class structure. Second, these embeddings are injected directly as constant membrane currents into leaky integrate-and-fire (LIF) neurons, removing the explicit neural encoding overhead. Third, each LIF neuron uses an adaptive dynamic threshold, driven by its membrane activity and recent spike history, bounded between [0.5, 2.0], which suppresses excessive firing. The network is trained end-to-end with surrogate gradients and backpropagation through time. On NSL-KDD, the proposed model reaches 93.35\% accuracy, a macro F1-score of 0.93, and an AUC of 0.966. Compared with a rate-encoded, fixed-threshold SNN baseline, it improves accuracy by 5.2\% points, lowers the false-positive rate from 8.4\% to 5.1\%, and nearly halves missed attacks (13.9\% to 7.6\%). These results show that direct feature injection combined with adaptive thresholding is a practical route to accurate, energy-aware intrusion detection on neuromorphic and edge hardware.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11367366/v1
Primary Topic
Network Security and Intrusion Detection
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Direct Current Feature Injection and Adaptive Dynamic Thresholding for Intrusion Detection using Leaky Integrate-and-Fire Neurons

Irshed Hussain, Vikash Kumar
Research Square
Network Security and Intrusion Detection
preprint

Direct Current Feature Injection and Adaptive Dynamic Thresholding for Intrusion Detection using Leaky Integrate-and-Fire Neurons

Irshed Hussain, Vikash Kumar
preprint en

Abstract

Abstract Intrusion detection systems deployed at the network edge should detect attacks accurately under severe class imbalance while operating within tight energy budgets. Spiking neural networks (SNNs) offer event-driven, low-power, and sparse computation. However, their accuracy is limited by three factors: lossy spike encoding of real-valued traffic features, uncontrolled firing under fixed thresholds, and poorly separable input representations. This paper proposes an SNN-based intrusion detection model that resolves the mentioned limitations. First, a class-aware discriminative autoencoder, trained with a staged schedule of reconstruction pre-training, decoder freezing, and class-weighted fine-tuning, maps PCA-reduced traffic features to compact embeddings that preserve minority-class structure. Second, these embeddings are injected directly as constant membrane currents into leaky integrate-and-fire (LIF) neurons, removing the explicit neural encoding overhead. Third, each LIF neuron uses an adaptive dynamic threshold, driven by its membrane activity and recent spike history, bounded between [0.5, 2.0], which suppresses excessive firing. The network is trained end-to-end with surrogate gradients and backpropagation through time. On NSL-KDD, the proposed model reaches 93.35\% accuracy, a macro F1-score of 0.93, and an AUC of 0.966. Compared with a rate-encoded, fixed-threshold SNN baseline, it improves accuracy by 5.2\% points, lowers the false-positive rate from 8.4\% to 5.1\%, and nearly halves missed attacks (13.9\% to 7.6\%). These results show that direct feature injection combined with adaptive thresholding is a practical route to accurate, energy-aware intrusion detection on neuromorphic and edge hardware.

Research Square
Malaviya National Institute of Technology Jaipur (IN), Sri Sivasubramaniya Nadar College of Engineering (IN)
Network Security and Intrusion Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon