Sex inference based on convolutional neural network analysis of fingerprint data.

The analysis of fingerprint features for inferring biological sex is a growing area of research in forensic science. This study presents a lightweight and well-validated convolutional neural network (CNN) as an alternative approach for this task. A dedicated dataset of 1,000 fingerprint images was collected from 200 volunteers (100 males and 100 females). To ensure rigorous evaluation of generalisation ability, an independent test set of 100 images from an additional 20 volunteers (10 males and 10 females) was held out for final assessment. The proposed CNN, featuring a dual-convolutional-layer architecture, was optimised using a cross-entropy loss function and the Adam optimiser. It achieved a validation accuracy of 91.00% and a test accuracy of 95.00%, with AUC values of 0.974 and 0.983, respectively. Supplementary fivefold cross-validation on the development cohort yielded a mean accuracy of 90.60% (SD: 2.04%), confirming stable performance. Class activation mapping (CAM) was employed to visualise the model's focus regions, enhancing interpretability and providing insights into biometric relevance. These results demonstrate that the model compares favourably with traditional methods, suggesting its potential as an efficient and reliable complementary tool for forensic identification.