Beyond Accuracy: Natural Adversarial Examples, Transferability and Robustness Metrics

Computer vision models can achieve high classification accuracy while remaining extremely sensitive to small changes in their inputs. This tutorial provides a practical and conceptual introduction to adversarial examples, ranging from deliberately generated perturbations produced by methods such as FGSM, PGD, and AutoAttack to so-called natural adversarial examples, which emerge without the intervention of an attacker as a consequence of acquisition noise or small variations between consecutive captures.

Drawing on experiments in microscopy, histopathology, object recognition, and architectural heritage classification, the tutorial will examine how the same perturbation may affect different neural network architectures in very different ways. In particular, we will introduce the concept of semi-universality, which measures how many models within an ensemble are fooled by each adversarial example.

The results show that an attack may achieve a very high ensemble attack success rate without simultaneously fooling every member of the ensemble. Under AutoAttack, the proportion of successful examples that fool all six evaluated models is approximately 22% for plankton-related datasets, 33% for architectural heritage images, and below 1% for histopathological datasets. These findings indicate that adversarial vulnerability depends not only on the model architecture, but also on the visual characteristics of the application domain.

The tutorial will also discuss how these phenomena should be evaluated beyond conventional clean accuracy. We will compare attack and robustness measures—including attack success rate, robust accuracy, cross-model transferability, leave-one-model-out evaluation, and semi-universality distributions—with distance-based and perceptual image-quality metrics such as L₂, SSIM, PSNR, VIF, IFC, and MAD.

The aim is to help participants jointly interpret attack effectiveness, perceptual distortion, and generalization across models, rather than reducing adversarial robustness to a single numerical score.

Generative AI for Visibility Enhancement in Challenging Driving Conditions: Night, Rain, and Fog

Poor visibility caused by nighttime conditions, rain, and fog significantly degrades image quality and affects reliable computer vision in intelligent transportation systems. This talk explores generative AI-based approaches for visibility enhancement, including GANs, image-to-image translation, and diffusion models. The discussion focuses on learning the transformation from degraded to clear images while preserving scene structure, texture, color, and illumination. Representative results for night enhancement, rain removal, and fog removal will be presented, along with commonly used reconstruction, perceptual, adversarial, and physics-informed losses and evaluation metrics. The talk will conclude with challenges and future directions toward unified, efficient generative models for all-weather image restoration and real-time vision applications.