Reproduce “Neuromorphic Cybersecurity with Semi-supervised Lifelong Learning” on OpenResearch

About this paper

Md Zesun Ahmed Mia, Malyaban Bal, Sen Lu, George M. Nishibuchi, Suhas Chelian, Srini Vasan, Abhronil Sengupta

The authors present a two-stage spiking neural network for intrusion detection: a fixed detector screens traffic, then a growing classifier processes flagged traffic. The classifier scales spike-timing-dependent plasticity updates with neuron-specific firing factors and can add or prune neurons. On UNSW-NB15, the dynamic system’s estimated accuracy is 85.3% versus 80.0% for a static baseline; the estimate has an arithmetic inconsistency.

Key insights

• A hierarchical SNN structure, mimicking biological processing, effectively manages computational resources by pre-filtering benign traffic before detailed attack classification.
• Dynamic structural plasticity (neuron growth and pruning) combined with Adaptive STDP (Ad-STDP) provides a bio-plausible mechanism for lifelong learning, enabling the SNN to adapt to new threats while retaining prior knowledge.
• The "firing factor" metric for neurons offers fine-grained control over synaptic plasticity, allowing newly added neurons to learn quickly and established neurons to stabilize their representations.

Summary by alphaXiv.