Using Convolutional Block Attention Module for Interpreting the Results of Artificial Neural Networks Operation in Handwritten Signature Identification

This study explores potential application of a Siamese Convolutional Neural Network (SNN) integrated with a Convolutional Block Attention Module (CBAM) in the field of handwritten signature identification. One well-known obstacle to the introduction of artificial neural networks into forensic expert practice is the low level of their interpretability. This limitation makes it difficult for experts to determine what patterns were identified by the neural network algorithm, and which ones of them form the basis of its forecast. Furthermore, in most contemporary studies on handwriting analysis, specialists tend to use the “classical” authorship attribution approach, treating the task as a particular case of classification. However, in authors’ view, this method often leads to Type II errors making the use of classification algorithms for identification purposes unacceptable. As an alternative, the authors propose to focus on the SNN architecture. To support these claims, a series of experiments were conducted as part of this study, demonstrating that modern mechanisms of attention – particularly CBAM – can partially interpret the neural network results. Additionally, the use of SNN helps to minimize the number of Type II errors compared to the traditional classification-based approach.