Count–accuracy coupling after ANN substrate transitions

Ascertain whether the empirical coupling between detection count reduction and mean Average Precision degradation under untargeted adversarial attacks persists after common substrate transitions of ANN object detectors, including compilation for edge accelerators and extreme low‑bit quantization (INT2).

Background

The paper introduces QCI to explicitly measure the coupling between detection count and accuracy and confirms this coupling on several ANN detectors under standard conditions and moderate quantization/pruning.

However, the authors note that more extreme or different substrate transitions (e.g., compilation for specific accelerators or INT2 quantization) have not been measured, and it remains unknown whether coupling would endure in those settings.

References

These measurements cannot be extrapolated beyond the conditions tested; substrate transitions that ANN detectors routinely undergo --- compilation for edge accelerators or extreme low-bit quantization (INT2) --- have not been measured, and whether count--accuracy coupling survives them is unknown.

— Fluently Lying: Adversarial Robustness Can Be Substrate-Dependent  (2604.00605 - Kang et al., 1 Apr 2026) in Section 6, Discussion, For object detection at large