Identification of Human vs AI Generated Handwriting Using Deep Learning Techniques

Advancements in AI possess a serious risk of mimicking human handwriting. Detecting whether handwriting is written by humans or generated by AI plays an important role in authentication, security and data validation. Robust systems are needed to differentiate between them. As human writing is being increasingly replicated by AI, we develop a strong classifier based on MobileNetV2 model with transfer learning trained and tested on a dataset of over 1000 human written and AI-generated handwriting. Proposed model's results are very accurate, more than DenseNet121 and VGG16 and it is also more efficient, achieves 99.75% accuracy, 99.50% precision, 100.00% recall, 99.75% F1-score, AUC 1.0000 tested on thousands of samples on training of 821.91 secs on 50 epochs. Handwriting verification tasks benefit best from the lightweight architecture and depthwise separable convolutions of MobileNetV2, as evidenced by its potential for real-world applications.