As AI models are increasingly deployed in real-world systems such as autonomous driving, intelligent transportation, and public safety, model evaluation is shifting from "accuracy on test sets" to "trustworthiness in open-world environments." Object detection models play a critical role in identifying and localizing surrounding objects, and their outputs are often used by downstream planning and decision-making modules. However, real-world conditions continually change in illumination, weather, occlusion, and scene composition. Even images that remain visually natural and semantically plausible may cause missed detections, false detections, or localization errors. The trustworthiness of an AI model therefore depends not only on high average accuracy, but also on its ability to make stable and consistent predictions under previously unseen yet plausible environmental variations.

In real-world applications, the opaque internal mechanisms and diverse architectures of object detection models make their latent failure modes difficult to uncover before deployment, creating unpredictable risks to system trustworthiness. Specifically, deployed models often operate as closed systems or online services, preventing testers from accessing model parameters or internal gradients. Meanwhile, different detection architectures rely on distinct mechanisms, including classification scores, region proposals, and bounding-box localization, making it difficult to characterize their failures using a unified approach. More importantly, trustworthiness testing must construct a special class of "counterfactual test scenarios": images that remain semantically plausible and visually consistent with the real world, yet cause clear detection errors. If test inputs contain conspicuous patches or noise, the observed failure may simply result from unrealistic artifacts; if the changes are overly conservative, they may fail to expose the brittle features on which the model actually relies. Balancing effective defect exposure with visual naturalness is therefore a central challenge in evaluating model trustworthiness in open-world environments.

Existing black-box testing methods primarily manipulate model outputs using fixed adversarial patches or pixel-level perturbations. These methods are often designed for a particular architecture or a single detection metric and struggle to balance testing effectiveness with visual naturalness. To address these limitations, the iSE Lab at Nanjing University, in collaboration with Nanjing University of Science and Technology, proposed UBA, a unified black-box adversarial testing approach. Rather than constructing perturbations around a single output metric, UBA jointly considers object classification, confidence, and localization to provide a unified characterization of failures across heterogeneous detection architectures. It then employs visualization-based analysis to identify the regions most influential to model predictions and contextually reconstructs these regions using surrounding scene information, generating high-value test cases in which "the scene remains plausible while the model fails." Experiments across eight mainstream object detection models and multiple real-world datasets show that UBA-generated samples exhibit strong cross-architecture transferability and expose model defects more effectively while preserving visual naturalness. Further training with these samples improves detection accuracy by 13.7 to 28.9 percentage points. This work moves adversarial testing beyond demonstrating failures under artificial perturbations toward uncovering vulnerabilities under semantically plausible changes, offering a new technical pathway toward a closed loop of risk discovery, vulnerability localization, and model enhancement.

These findings were published in the paper "UBA: A Unified Black-Box Adversarial Testing for Object Detection via Visualization-Based Contextual Reconstruction" in the International Journal of Computer Vision (IJCV), as part of the special issue Ensuring Trustworthiness in Open-World Visual Recognition. Dr. Jiawei Liu, the paper’s first author, received her Ph.D. from the School of Software at Nanjing University in 2025. Her research focuses on quality assurance for AI models and operators. Building on earlier work in AI model testing, this study further integrates interpretability-guided analysis with contextual reconstruction, providing an important reference for advancing object detection trustworthiness testing from exposing failures under artificial perturbations to discovering risks in semantically plausible real-world scenarios and strengthening models accordingly.