Abstract: To address the difficulty of existing specification generation methods in simultaneously covering heterogeneous safety requirements, including functional safety, cybersecurity, and safety of the intended functionality, under the cyber attacks in dynamic traffic environments, we propose a specification generation framework for autonomous driving systems. Firstly, we construct a unified signal temporal logic (STL) semantic space based on heterogeneous requirements, providing a formal semantic foundation for specification generation. Secondly, we combine parameterized scenario modeling and attack-injection mechanisms to construct autonomous driving scenarios that cover both nominal and attack-induced behaviors, thereby forming a structured trajectory space for specification generation. Finally, we design a Transformer-based collaborative structure-parameter learning model, in which a dual-branch inference mechanism is employed to generate STL specifications. Results from high-fidelity simulation and physical experiments show that the proposed approach can stably generate STL specifications with clear physical semantics across diverse driving scenarios and attack conditions, with a structure prediction accuracy of 89.1% and prediction accuracies above 91.0% for critical parameters, such as safe distance and time-to-collision (TTC).