Multiscale Astrocyte Network Calcium Dynamics for Biologically-Informed Intelligence in Anomaly Detection

Network anomaly detectors trained offline can lose performance as traffic patterns and attack distributions change over time. We introduce an astrocyte-informed Ca²⁺-modulated learning framework that couples a multicellular Ca²⁺ simulator to a deep neural network (DNN). The simulator captures IP3-mediated Ca²⁺ release, SERCA uptake, and conductance-aware diffusion through gap junctions. Its output contributes to a metaplastic learning rule in which local, heterosynaptic, postsynaptic, supervisory, and multicellular Ca²⁺ signals regulate gradient updates through an adaptive threshold. We evaluate the framework on CTU-13 Neris traffic using conventional and leakage-controlled chronological settings, strong adaptive DNN baselines, and five-seed ablations. Under temporal distribution shift, the full Ca²⁺ model gives the strongest mean performance among the evaluated adaptive baselines and reduces missed detections relative to the matched DNN. Factorial ablation identifies Ca²⁺ gating as the main contributor to this gain: Laplacian regularisation alone provides little improvement, whereas removing the multicellular Ca²⁺ term substantially reduces performance. Randomised, shuffled, and constant-signal controls further show that the benefit is linked to Ca²⁺-dependent metaplastic regulation, while the precise temporal waveform is not uniquely required. These results support astrocyte-informed Ca²⁺ signalling as a mechanism for regulating neural-network adaptation under changing data distributions.