In an ordinary neural network, every neuron passes a number to the next layer on every pass. In a spiking neural network, a neuron accumulates input until it crosses a threshold, fires a spike, and resets. Information is carried by which neurons fire and when. This is the model neuromorphic chips are built to run, and it suits event-based sensors that produce streams of changes instead of full frames.
Because timing is part of the computation, the same inputs in a different rhythm can give a different answer. That makes spiking networks efficient, and it also gives attackers new levers: research has shown backdoors triggered by flashing light, and power and electromagnetic measurements that leak the network’s parameters.
Do not assume defences built for image classifiers carry across. Several studies found that pruning, fine-tuning and image-domain backdoor defences often failed against attacks designed for spiking networks and neuromorphic data.