Fingerprint Presentation Attack Detection: A Comparative Evaluation of Support Vector Machine and K-Nearest Neighbor Classification

Abstract

The human hand consists of the wrist, palm, and fingers. It is used to feel, hold, and move objects. Each finger has a unique ridge pattern that develops before birth and remains unchanged throughout a person’s lifetime, which is a major reason why it is often used for identification purposes. Fingerprint recognition is used to verify an identity by comparing a person’s fingerprint with a previously registered fingerprint. Despite its widespread use, fingerprint recognition remains vulnerable to presentation attacks. A presentation attack occurs when a fake biometric sample is presented to a sensor to impersonate a genuine user. This study compares Support Vector Machine and K-Nearest Neighbor for detecting fake fingerprints using Local Binary Pattern (LBP) features from the FVC2000 and SOCOfing datasets. On the combined SOCOFing dataset, SVM achieved an accuracy of 86.8%, a precision of 71.4%, a recall of 87.1%, and an F1-score of 75.6%. Both models showed improved performance from the Easy to the Hard tier, while performance on the combined tier was lower. SVM had a peak accuracy of 97.6% and ROC-AUC of 99.7% on the most heavily altered (Hard) subset alone. This suggests that, counterintuitively, heavier digital alteration may be easier to detect than subtle alteration. On the FVC2000 benchmark, both classifiers exceeded 99.5% accuracy, although this result may reflect the relative ease of separating genuine sensor captures from synthetically generated samples. These results indicate that classical machine learning models remain a viable and competitive option for fingerprint presentation attack detection, with SVM outperforming KNN, while detection performance is strongly dependent on alteration or attack severity rather than being uniform across difficulty levels.