Robustness Verification of Classification Through Image Perturbations

Well-trained deep neural networks consistently show a strong level of confidence in their predictions, but their robustness in real-world scenarios is still a cause for concern and various trustability and reliability studies for the same have been conducted. We introduce a framework for evaluating model robustness by estimating weighted confidence, specifically examining pretrained ResNet and Swin Transformer architectures. By integrating image quality assessments with weighted perturbation analysis which leverages the Generalized Normal Distribution, we generate weighted confidence scores that highlight notable discrepancies and failures in architectural robustness. In our findings, we show that this method can pinpoint reliability shortcomings in highly confident predictions and unveil unique behavior in the way architecture reacts to perturbations, fostering the advancement of resilient AI systems.