Damage Resistance of an fMRI-Spiking Neural Network Based on Speech Recognition Against Stochastic Attack

Brain-like models are commonly used for pattern recognition, but they face significant performance degradation in neuromorphic hardware when exposed to complex electromagnetic environments. The human brain has adaptability to the exterior attack, and we expect that incorporating bio-plausibility into a brain-like model will enhance its robustness. However, brain-like models currently lack bio-plausibility. Therefore, we construct a spiking neural network (SNN) whose topology is constrained by human brain functional Magnetic Resonance Imaging (fMRI), called fMRI-SNN. To certify its damage resistance, we investigate speech recognition accuracy against stochastic attack. To reveal its damage-resistant mechanism, we explore the neural electrical features, adaptive modulation of synaptic plasticity, and topological features against stochastic attack. Research shows that fMRI-SNN surpasses SNNs with distinct topologies in recognition accuracy against stochastic attack, notably maintaining similar accuracy levels before and after stochastic attacks when the damage proportion is below 30%, demonstrating that our method improves the damage resistance of brain-like models. In addition, the change in neural electrical activity serves as interior manifestation, corresponding to the damage resistance of SNNs for recognition tasks, while the synaptic plasticity serves as the inherent determinant of the damage resistance, and the topology serves as a determinant impacting the damage resistance.