Robustness Assessment for Image Classification Systems to Combined Common Corruptions Based on Decision Boundary Scenarios

Assessing the robustness of image classification systems against common corruptions is crucial for enhancing their reliability in real-world applications. Current methods predominantly assess the robustness of image classification systems by their classification accuracy on the corrupted dataset with fixed corruption severity levels, while neglecting the compounded effects of multiple corruptions and unique characteristics of systems. This paper proposes a framework, RealBoundary, to quantify the robustness of image classification systems to combined corruptions by exploring their decision boundaries. We focus on combined corruptions and introduce an efficient adaptive searching algorithm to search for decision boundary scenarios within the corruption severity parameter space. We define two quantitative robustness metrics, the tightest robustness (TR) and the statistical robustness (SR), considering the closest and average distance from the clean images to the decision boundary. Experimental results demonstrate several interesting findings. The robustness of the classification system under different images displays an approximate proportional relation to image classification confidence scores. Additionally, under combined corruptions, smaller image degradation could cause the system to lose robustness. Under the same corruption, the decision boundaries of the system share similar geometric landscapes across different input images.