This paper introduces an adversarial training framework that enhances the robustness of deep hedging strategies against distributional shifts, demonstrating superior out-of-sample performance and resilience compared to classical models.