Bridging the Gap Between AI and Reality • Rhodes, Greece
Time: Tuesday, 4.11
Room: Room C
Authors: Elena Hoemann
Abstract: AI-based systems are increasingly used in daily life, yet their adoption in safety-critical domains such as automated driving remains challenging due to insufficient safety assurance. Since conventional verification methods cannot be applied to black-box AI models, new safety methodologies are required. This study investigates the role of data engineering as a key component of AI safety, focusing on how training data distribution affects model accuracy. Experiments were conducted using the MNIST dataset and extended to the GTSRB traffic sign dataset to evaluate transferability to automotive applications. Results show that dataset size and class complexity significantly influence accuracy, while distribution imbalances have a smaller effect. Notably, reducing samples of safety-critical classes, such as stop signs, leads to performance degradation. These findings highlight the importance of data-centric approaches for developing safe, reliable AI in automated driving systems.