Problem

The increasing realism of synthetic fingerprint images, driven by advancements in generative artificial intelligence (GenAI), poses a significant threat to fingerprint recognition systems through data injection attacks. While deep neural networks (DNNs) can be trained to detect synthetic images, they often fail to generalize to synthetic fingerprints generated by unseen GenAI models. This leads to a catastrophic collapse in detection accuracy when encountering new, evolving synthetic styles, highlighting the need for adaptive detection methods.

Method

The authors formulate synthetic fingerprint detection as a continual few-shot adaptation problem. Their DNN-based framework consists of a pre-processor, a pre-trained DNN backbone as a feature extractor (e.g., ViT-Small), and a classifier head (MLP). To learn the base detector, binary cross-entropy (CE) loss is minimized. For continual few-shot adaptation to new, unseen synthetic styles, two main strategies are employed to mitigate catastrophic forgetting: (1) Experience Replay: A small number of samples from previously known styles are retained in a memory buffer and replayed during fine-tuning alongside samples from the new style. (2) Combined Loss Function: Adaptation uses a combination of binary cross-entropy loss and supervised contrastive (SC) loss. The SC loss helps pull representations of the same class closer while pushing different classes apart, stabilizing representation learning with limited new data. The model parameters are initialized from the previous adaptation step.

Results

Experiments using various DNN backbones and diverse real and synthetic fingerprint datasets demonstrated several key findings. The base detector showed very poor zero-shot generalization, with its True Detection Rate (TDR) collapsing significantly when faced with unseen synthetic styles. Few-shot adaptation proved beneficial, substantially improving detection performance on new styles, though the gains decreased with more sophisticated GAN-generated images. The proposed ViT-CE+SC model consistently outperformed ViT-CE (using only CE loss) across all adaptation steps, indicating the effectiveness of the supervised contrastive loss in stabilizing learning with small sample sizes. Experience replay was crucial in minimizing catastrophic forgetting, preserving performance on previously learned styles, and even showing some backward knowledge transfer. The performance gains from increasing adaptation sample size were most significant from 10 to 50 samples, with diminishing returns thereafter. The approach was also shown to be backbone-agnostic, with more advanced backbones further improving accuracy.

Implications

This work presents the first study on synthetic fingerprint detection, framing it as a continual few-shot adaptation challenge. The proposed methodology, combining binary cross-entropy and supervised contrastive losses with experience replay, offers a robust solution for dynamically evolving synthetic fingerprint detectors. By effectively balancing rapid adaptation to novel synthetic styles and retaining knowledge of previously learned styles, this approach holds significant potential as a crucial defense mechanism against sophisticated data injection attacks in fingerprint recognition systems.