Testing Techniques in Deep Learning Systems: A Survey

---

Résumé

Deep learning (DL) systems have achieved remarkable success across domains such as computer vision, natural language processing, and autonomous driving. However, their deployment in safety-critical and security-sensitive scenarios necessitates thorough testing to ensure reliability and trustworthiness. Unlike traditional software, DL models lack well-defined logic and rely heavily on data, introducing unique challenges such as the oracle problem, high labeling cost, and bias detection. Although early survey has provided high-level overviews of DL testing, the field has evolved significantly over the past five years, with a surge of new techniques and a growing emphasis on practical concerns like test efficiency and fairness. This paper presents a comprehensive survey and analysis of 236 DL testing studies, covering six major research communities. We categorize the existing techniques based on six quality properties: efficiency, robustness & security, fairness, interpretability, correctness, and privacy. We provide in-depth analyses of the papers in each category using comparative analyses, distributional trends, and temporal evolution, obtaining 14 core findings. Furthermore, we highlight the emerging frontier of testing for DL, presenting recent trends, unique challenges, and promising future directions.  

Mots-clés

Aucun mot-clé renseigné.