SNN-IDS: Deploying SNN-Equivalent Intrusion Detection on a Commodity MCU NPU

We deploy a spiking neural network (SNN)-equivalent intrusion detection system (IDS) on the STM32N6570-DK, a commodity ARM Cortex-M55 MCU with the Neural-ART NPU. Exploiting the approximate equivalence between single-timestep (T=1) SNN inference and INT8 quantized ANN inference, we compile a lightweight MLP classifier to the NPU without neuromorphic hardware. Evaluated on NSL-KDD (5-class) and UNSW-NB15 (10-class) with 10 random seeds, the ReLU model achieves 78.86±1.32% and 64.75±0.61% overall accuracy, respectively. INT8 accuracy stays within 1 percentage point of FP32 across all 24 tested calibration configurations, and layer-wise analysis shows 99.0% final prediction agreement between FP32 and INT8 models. On the NPU, the INT8 model infers in 0.46 ms on NSL-KDD and 0.29 ms on UNSW-NB15 (100% NPU execution), occupying 120.6–137.7 KB Flash and 0.5–1.25 KB RAM. A comparison with QCFS activation reveals that the Floor operator falls back to CPU on this NPU, adding 17.6% latency. Tree-based baselines (Random Forest, XGBoost) confirm that the MLP offers the best accuracy on NSL-KDD while being the only model eligible for NPU acceleration. To our knowledge, this is the first IDS deployment on an ARM Cortex-M NPU and the first empirical validation of T=1 SNN–ANN equivalence on commercial NPU silicon.