APSS Metrics for Fault Detection: Area, Position, Symmetry, and Shape in Image Segmentation

In recent years, the impact of hardware-induced faults on neural networks performing image classification tasks has gained a lot of attention. Specifically, failures have been directly associated with wrong classifications. When it comes to different tasks, this association is less explicit. For example, the critical impact of hardware-induced faults on image segmentation tasks is less interpretable. In this work, we propose a novel technique for detecting critical permanent faults, relying on a dataset profiling phase to extract four metrics. These metrics are designed to assess and monitor the area, position, symmetry, and shape of prediction patterns across the output mask at the pixel level. Validation was performed through a statistical fault injection campaign on Fast-SCNN model trained on Cityscapes. To evaluate the effectiveness of the proposed method, a Faulty Output Dataset (FOD) was developed and employed to compare state-of-the-art (SOTA) metrics, such as Pixel Accuracy (PA) and mean Intersection over Union (mIoU), with the proposed one. The results show a high capability to detect critical faults, with an accuracy greater than 99\%, comparable to SOTA methods, but with the advantage that the proposed method does not require a golden mask, increasing its applicability in real-world scenarios.