Lifelong iris presentation attack detection without forgetting

Despite the promising results achieved by deep iris presentation attack detection (PAD) in dataset-specific scenarios, the advanced approach remains vulnerable to novel attacks. Real-world attacks evolve over time. Typically, fine-tuning and retraining from scratch are employed to incrementally learn new attacks. However, fine-tuning degrades performance on old attacks, i.e., catastrophic forgetting. Retraining on all data is unavailable due to data privacy. To address these issues, we are the first to propose a lifelong iris PAD to incrementally learn new attacks without storing old data. Our approach utilizes a prompt pool to preserve attack-independent and attack-shared knowledge, wherein learnable prompts aid in prediction by the pre-trained Vision Transformer (ViT). Furthermore, adaptive attention masks for sequential new attacks are applied to pre-trained ViT. Consequently, our method improves plasticity while preserving stability. Extensive experiments are performed on our building dataset combing IITD and CASIA to evaluate iris PAD in incremental learning. Our proposed method obtains competitive performance over state-of-the-art Iris PAD schemes.