Causality of hardware constraints for Quality Corruption

Determine whether the conjunction of the three neuromorphic hardware constraints—binary spike communication, accumulate-only arithmetic, and the absence of dense matrix computation—causally induces the Quality Corruption failure mode (i.e., count‑preserving accuracy collapse under untargeted PGD attacks) in hardware‑deployable spiking neural network object detectors such as EMS‑YOLO.

Background

The paper finds that under identical untargeted PGD attacks, three non‑deployable SNN detectors and their ANN baselines consistently exhibit suppression, whereas EMS‑YOLO—the only detector satisfying all three neuromorphic hardware constraints—exhibits Quality Corruption, where detection counts remain while mAP collapses. This pattern persists across norms and losses.

A membrane‑targeted probe (FMP) suggests that membrane dynamics are exploitable exclusively in the hardware‑deployable pipeline, but the authors emphasize that they cannot infer causality from a single model and set of constraints. Hence, establishing whether these constraints are causally responsible for Quality Corruption remains an open question.