Evidential Value of a Writer-Independent CNN-LSTM\HOG-LSTM Model with Overfitting Control for Offline Signature Verification in Criminal Cases

Signature forgery is of great importance in both forensic science and criminal law due to its significance in establishing the identities and legal personality of individuals. However, the process of determining legal personality relies on the visual and other senses to distinguish individuals and therefore may involve subjective interpretations. Offline signature verification techniques, on the other hand, utilize more objective methods based on numerical data and indistinguishable to the human eye. To generate numerical results for signature forgery, Gradient Histogram (HOG) and ResNet-18 (CNN) were used separately, and feature matrices were extracted within this framework. A five-layer, author-independent cross-validation protocol was employed; in this protocol, approximately 64% of the authors in each layer were allocated for training, 16% for validation, and 20% for testing, and a Long Short-Term Memory (LSTM) network was used for classification. The performance of the model was evaluated with the BHSig260 dataset from 160 Hindi and 100 Bengali individuals, each containing 24 genuine and 30 forged signatures, and the same parameters were used as both CEDAR and BHSig260. The CEDAR dataset was examined in detail because it uses the Latin alphabet, but the numerical performance of the model in Bengali and Hindi was also considered. At this point, the CEDAR offline signature dataset was used to visualize the results obtained by the model. This dataset contains 1320 genuine signatures from 55 real signatories and 1320 forged signatures corresponding to these genuine signatures. Numerical evaluations and graphical outputs measuring different criteria were obtained with this dataset, and the consistency of the numerical results was evaluated. In this context, calibration and DET analyses measured the reliability of the results obtained in the confusion matrix, and the model's relationship with perfection and hesitant-bold behavior was also observed. In light of these evaluations, this study examines the acceptability and reliability of AI-based signature verification systems in criminal proceedings and discusses their contributions to forensic cases. Based on this, it is possible to use all outputs obtained from AI-based signature verification techniques as a decision support tool in determining guilt.