When Pixels Lie: The Hidden Toll of Concept Drift
Abstract
Concept drift poses a major challenge to the reliability and transparency of computer vision systems. This study provides a systematic analysis of how four drift types, such as noise, blur, rotation, and brightness, affect both model performance and Grad-CAM/Grad-CAM++ explainability across CIFAR-10 and Tiny ImageNet datasets using three CNN architectures (ResNet-18, DenseNet-121, ShuffleNet v2). Following recent robustness literature, we treat these controlled test-time transformations as synthetic covariate-shift proxies for the input-distribution component of drift rather than as temporal concept drift in the strict sense. For CIFAR-10, accuracy declines from a baseline of 79–86% to as low as 24% under strong noise drift, while SSIM and IoU between baseline and drifted heatmaps fall from nearly 100% to as low as 18% and 25%, respectively. On Tiny ImageNet, the degradation is even more pronounced: top-1 accuracy collapses from 40–60% to below 4% under the strongest noise drift, with SSIM and IoU dropping to as low as 10% (per-architecture values are reported in the Results and Discussion section). Rotation and brightness prove markedly milder than noise and blur, indicating that high-frequency corruptions are the principal threat to both accuracy and attribution stability. Image-level paired tests (Wilcoxon signed-rank and paired t-test) with effect sizes and bootstrap confidence intervals, complemented by pixel-level distributional diagnostics (KS, Anderson-Darling, Wilcoxon rank-sum), confirm significant shifts in heatmap distributions, revealing that drift alters not only model outputs but also internal attribution patterns. The results demonstrate that different drift mechanisms degrade interpretability in distinct ways and that explainability metrics are often more sensitive to drift than accuracy itself. This work provides, to the best of our knowledge, one of the first unified, quantitative evaluations of explainability degradation under controlled visual drift, highlighting the need for adaptive, drift-aware XAI pipelines in dynamic real-world environments.