Fuzzing AI Systems: Foundations, Techniques, and Open Challenges

Engineering notes

Our analysis shows rapid but uneven growth, strong target-dependence in technique design, persistent oracle-construction challenges, failures beyond crashes, and the need for reproducible benchmarks and cross-layer evaluation.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Artificial Intelligence (AI)-enabled systems now appear in settings where failures can affect software quality, safety, and security, including autonomous driving, software engineering tools, and large language model services. Fuzzing offers a practical way to stress these systems by generating or mutating test inputs, but the resulting research is scattered across different targets, techniques, oracle designs, and failure definitions. This survey synthesizes 125 primary studies and organizes the literature using a taxonomy of testing targets, system layers, technique families, input-generation and mutation strategies, oracles, and failure types. Our analysis shows rapid but uneven growth, strong target-dependence in technique design, persistent oracle-construction challenges, failures beyond crashes, and the need for reproducible benchmarks and cross-layer evaluation.

Links and sources

Comments