Performance Analysis of an Adapted ResNet-50 Model for Fingerprint Recognition on Synthetically Modified Images

Abstract

Fingerprint recognition systems are widely deployed in security and forensic applications, yet their ability to identify individuals
whose fingerprints have been intentionally or naturally altered remains a practical concern. This paper evaluates the performance of a ResNet-50 model adapted for fingerprint identity classification using the Sokoto Coventry Fingerprint (SOCOFing) dataset, which provides 6,000 original images from 600 individuals alongside 49,270 synthetically modified images generated via the STRANGE toolbox. The model was trained exclusively on original fingerprint images, reflecting the realistic deployment scenario where operational databases contain only unaltered enrollment records. Six validation sets were constructed: three organized by modification difficulty level (easy, medium, hard) and three by modification type (obliteration, Z-cut, central rotation). Training used a transfer learning approach with full fine-tuning, the Adam optimizer, and Sparse Categorical Cross-Entropy loss over configurations of 20 and 50 training epochs. Results show that validation accuracy at 50 epochs decreased from 94.98% for easy modifications to 73.88% for hard modifications, confirming a consistent relationship between alteration severity and classification difficulty. Among modification types, obliteration and central rotation produced stable training dynamics and accuracy above 91% at 50 epochs, while Z-cut modifications caused repeated episodes of validation accuracy collapse when training extended beyond 20 epochs. At 20 epochs, Z-cut validation accuracy reached 75.13%, the highest among all modification types at that epoch count, indicating that the model generalizes quickly to Z-cut images but subsequently overfits. These results demonstrate that ResNet-50 trained only on unmodified fingerprints retains useful recognition ability across most alteration scenarios, while identifying Z-cut as a source of generalization instability that warrants targeted mitigation strategies such as early stopping and regularization.