Table 6 Accuracy improvement and significance: summary of the overall accuracy (mean of single-failure cases) for baseline versus suggested technique, the absolute improvement (recovery margin), and the related p value from a paired t test (10 runs). The relative reduction in mistake rate is also provided in parentheses for context. All improvements are statistically significant (p < 0.05).
From: Modifier guided resilient CNN inference enables fault-tolerant edge collaboration for IoT
Dataset and fusion | Baseline accuracy | Proposed accuracy | Improvement (abs) | p value (sig.) | Error reduction (%) |
|---|---|---|---|---|---|
MNIST—Max Pool | 81.11% ± 0.5 | 82.60% ± 0.4 | + 1.49% | p = 0.0021 (signif.) | ~ 7.9% reduction |
MNIST—Avg Pool | 93.37% ± 0.3 | 94.28% ± 0.2 | + 0.91% | p = 0.005 (signif.) | ~ 13.7% reduction |
CIFAR-10—Max Pool | 70.57% ± 1.0 | 71.51% ± 0.9 | + 0.94% | p = 0.016 (signif.) | ~ 3.2% reduction |
CIFAR-10—Avg Pool | 79.53% ± 0.8 | 80.16% ± 0.7 | + 0.63% | p = 0.027 (signif.) | ~ 3.0% reduction |
