ForgeXplain: An Explainable Machine Learning System for Offline Signature Forgery Detection

Achieving accurate signature verification is critical in many contexts including financial transactions and legal contracts, for reasons of security and fraud risk. The authors describe an offline signature verification system called ForgeXplain that uses hand engineered image features in conjunction with machine learning algorithms and explainable AI. From pre-processed signature images the system extracts 23 hand designed geometric and intensity features that in turn feed into support vector machine (SVM) and random forest (RF) models in order to predict fake from real signatures, with SHAP values explaining the decision on a feature-by-feature basis. A test of 1,320 real and 1,320 fake signatures from CEDAR offline found that Random Forest was correct 91.5 percent of the time and achieved a ROC-AUC score of 0.966, better than SVM. This shows that ForgeXplain can tackle off-line signature forgery detection effectively and faithfully.