Sponge Attacks Increase SNN Inference Energy by up to 2.6x
July 31, 2026
Gradient-based optimization can craft adversarial spike trains for Spiking Neural Networks that inflate synaptic operations (SynOps) by 1.5-2.6x. These sponge attacks bypass correctness-based monitoring by maintaining accuracy while significantly increasing energy consumption on neuromorphic hardware.
HOW THIS AFFECTS YOU
●
builderEdge deployment of SNNs requires monitoring for unexpected spikes in power consumption to prevent battery exhaustion attacks.
●
researcherYou must account for energy-based adversarial threats when evaluating SNN security.