ASTDP-GAD framework achieves energy-efficient anomaly detection with spiking neural nets

A new paper from Abdul Joseph Fofanah and colleagues introduces ASTDP-GAD, an adaptive spiking temporal dynamics plasticity framework that tackles anomaly detection in dynamic networks with unprecedented energy efficiency. The system fuses spiking neural networks (SNNs) with spike-timing-dependent plasticity (STDP) and graph-based learning, enabling it to detect irregularities in real-time streaming data while consuming orders of magnitude less power than traditional deep learning approaches.

The framework comprises six key innovations: temporal spike graph encoding with adaptive Leaky Integrate-and-Fire (LIF) dynamics that preserve input features linearly with simulation steps; LIF-based graph attention with lateral inhibition that can approximate any continuous attention function; event-driven hypergraph memory using STDP-inspired prototype updates that converge to optimal representations; spike rate contrast pooling that provably selects anomalous nodes; adaptive STDP layers capturing causal temporal relationships with stable convergence; and multi-scale temporal convolution combined with multi-factor anomaly fusion that reduces output variance by up to 5x.

Extensive experiments across nine datasets on both dynamic and static graphs demonstrate that ASTDP-GAD matches or exceeds state-of-the-art anomaly detection accuracy while maintaining biological plausibility and suitability for neuromorphic deployment. The framework's theoretical guarantees—from information preservation to convergence bounds—make it a compelling candidate for edge applications where power budgets are tight, such as cybersecurity, industrial IoT monitoring, and financial fraud detection.