Hybrid siamese architecture for detecting genuine training data fingerprints in medical generative adversarial networks

The rapid growth in generative models for medical imaging demands robust methodologies for preserving patient privacy and ensuring that the synthetic data remains authentic. The leaking of identifiable patterns from real clinical data into generated images presents serious difficulties in terms of regulatory compliance, ethical deployment, and clinical trust. This work, therefore, attempts to identify the actual clinical images mistakenly included in a generative adversarial network during its training and enables privacy-preserving generative modeling. A hybrid Siamese architecture consisting of ResNet-50 and vision transformer backbones is introduced to learn robust feature representations. Cross-attention is incorporated into the model for multiscale feature fusion, and an adaptive similarity metric is developed to improve discriminative comparison of paired images. We use a hybrid loss function that combines an angular term and a binary cross-entropy term during training for stable optimization. Through extensive evaluation on lung CT scan datasets, we show that our model has shown the highest average accuracy ([Formula: see text]) and an equally good F1-score ([Formula: see text]) This is a reasonable classification performance and also points out the difficulties involved in the detection of generative fingerprints. These findings support the superior discriminative power of the model for subtle generative 'fingerprints'. With a presented novel privacy-preserving framework for generative model validation, the proposed method can be a contribution to the development of secure and ethical validation for clinical AI systems.