A hybrid spiking convolutional neural framework with extreme learning machine for enhanced anomaly detection in...

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Computational Neuroscience

Background:

• Real-time cybersecurity systems struggle with detecting breaches in dynamic, unbalanced network data.
• Existing methods often lack efficiency and adaptability for streaming network traffic.

Purpose of the Study:

• To develop a unified hybrid framework (SCNN-OeSNN-ELM) for enhanced real-time network breach detection.
• To improve spatiotemporal feature extraction and anomaly detection accuracy in network streams.

Main Methods:

• Utilized Gaussian Receptive Fields (GRFs) to encode streaming inputs into spike trains.
• Employed spiking convolutional layers for spatiotemporal feature extraction.
• Integrated an online evolving spiking neural network (OeSNN-UAD) with an extreme learning machine (ELM) for optimized learning and detection.

Main Results:

• The SCNN-OeSNN-ELM framework demonstrated superior performance over existing spiking-based baselines on NAB and CIC-IDS2017 datasets.
• Achieved relative detection increases of 3-10% on NAB and 3-6% on CIC-IDS2017.
• Showcased statistically significant improvements in precision, recall, F1-score, balanced accuracy, and MCC, alongside enhanced latency and energy efficiency.

Conclusions:

• The proposed SCNN-OeSNN-ELM framework offers an effective solution for real-time network anomaly detection.
• The hybrid model provides low-memory operation, online adaptation, and fast, non-iterative learning.
• The framework's event-driven nature and analytical optimization contribute to its efficiency and performance gains.