Sylvester Kaczmarek (Imperial College London)
Static defenses are brittle against the non-stationary threats common in long-duration space missions. We propose a framework for self-organizing resilience where a Spiking Neural Network (SNN) dynamically adapts its own structure to counter novel adversarial tactics. Governed by an informationtheoretic objective that balances representational fidelity against computational cost, the network autonomously grows or prunes neural populations to specialize against previously unseen threat signatures. We present preliminary results from a cislunar gateway case study where the adaptive SNN is subjected to a low-rate data injection attack designed to evade static detectors. The adaptive model successfully learned the new threat pattern, reducing per-window inference time by over 40% compared to its static counterpart, with no degradation in nominal performance. We provide explicit triggers, a two-stage commit with rollback, and an audit log, treating online adaptation as a security control bounded by runtime envelopes.