Abstract: Iris recognition, due to its uniqueness, stability, and high accuracy, has been widely applied in fields such as financial payments and public security. However, with continuous technological advancements, attackers are attempting to forge iris features in various ways to bypass identity verification systems. Iris liveness detection aims to distinguish bona fide iris samples from various types of attack iris images. Existing methods for iris liveness detection often suffer from inadequate robustness and generalization when faced with complex environmental conditions and diverse attack techniques, making them insufficient for practical applications. To address the limitation of existing iris liveness detection methods, which rely solely on visual feature modeling and exhibit insufficient generalization in scenarios involving both physical and digital attacks, this paper proposes an iris liveness detection method that integrates a text-driven attention mechanism with the vision foundation model, named Text-Guided Attention for IPAD (TGA-IPAD), which fuses information from both visual and textual modalities, leveraging the rich semantic knowledge of pre-trained language models to guide separation of bona fide and attack iris samples in the representation space. By utilizing textual semantics as a form of weak supervision, the model's learning process is explicitly constrained and guided, leading to more robust and discriminative iris feature representations. The architecture is designed to overcome the challenges of adapting generic vision-language models to the fine-grained domain of iris liveness detection. Specifically, to enhance the method’s ability to perceive iris features, a fine-grained visual-text alignment framework based on textual prompts is introduced, enabling precise matching between iris image features and semantic information, thereby directing the model's attention to subtle, attack-relevant regions such as texture irregularities, printing artifacts and so on. What’s more, multiple learnable textual prompts are employed to avoid bias from a single description, and their mean prototype features serve as the semantic anchor for alignment. Furthermore, the method employs text-conditioned constraints based on slot attention to decouple bona fide and spoof features in the latent space, thereby improving generalization across diverse attack types. Experimental results on the challenging LivDet-Iris 2023 competition datasets demonstrate the superiority of the proposed method over the winner methods from these competitions and several state-of-the-art (SOTA) approaches. Specifically, compared to the best SOTA method, the overall average classification error rate, ACER1 and ACER2, are reduced by 12.03% and 6.46%, respectively. Moreover, for synthetic iris attacks and physical attacks, the attack presentation classification error rate (APCER) are reduced by 27.98% and 1.43%, validating its effectiveness in open environment. Beyond quantitative evaluations, qualitative experiments utilizing textual prompts demonstrate that the model can effectively capture discriminative features specific to bona fide and attack irises, confirming that the integration of textual guidance leads to a more generalizable understanding of iris liveness detection. In conclusion, this study not only offers a high-performance solution for iris liveness detection but also pioneers a promising technical direction by effectively harnessing vision-language foundational models to significantly enhance the security and generalizability of biometric systems.