A Unique Approach to Efficient Fraudulent Signature Detection Using Deep Convolutional Neural Network, Xception, and EfficientNet
Srishti Lodha, Harsh Malani · 2022
Signatures act as the key to the validation of important documents, both in offline and digitized forms. This also means that forged signatures are vastly used to commit fraud. With Machine Learning gaining importance in every field, manual techniques are becoming obsolete and inefficient in several technology-related areas, including fraudulent signature detection. Hence, this research focuses on the use of Deep Learning to construct a generalized yet accurate model, which can directly classify a given signature as genuine or forged instead of having a separate class for genuine and fraud for each of the signatures in the output layer. We train separate models with Deep Convolutional Neural Network, Xception, and EfficientNet respectively, then, we analyze and compare the results obtained by these models, as well as by the existing related works, inclusive of various performance metrics. After rigorous hyperparameter tuning and experimenting, we obtained the best results from Xception, with the highest training accuracy of 100% and validation accuracy of 98.02%.