The verification of handwritten signatures is a significant area of research in computer vision and machine learning (ML). Handwritten signatures serve as unique biometric identifiers, making it essential to distinguish between genuine and forged signatures. This binary classification task is crucial in legal and financial contexts to prevent fraud and protect customers from potential losses. However, verifying offline handwritten signatures is challenging due to variations in handwriting influenced by factors such as mood, fatigue, writing surface, and writing instrument. This research paper focuses on recognizing offline handwritten Arabic signatures using deep learning (DL), specifically transfer learning (TL) technique which is called “Inception-V3 TL model”. Three distinct datasets are used to build a model for recognizing signatures. The first dataset is referred to as Dataset1. It is an English signature dataset called "CEDAR" which contains 1,320 genuine and 1,320 forged signatures. Dataset1 is publicly available at: https://www.kaggle.com/datasets/shreelakshmigp/cedardataset .The second dataset is referred to as Dataset2. It is a new Arabic signature dataset created for this research which contains 1,320 genuine and 1,320 forged signatures. The third dataset is referred to as Dataset3. It is created by merging the English and Arabic signature datasets (Dataset1 and Dataset2). The Inception-V3 TL model is trained on these distinct datasets (Dataset1, Dataset2, and Dataset3). Both normal training and k-fold cross-validation (CV) methods are applied to evaluate the model’s performance, ensuring robustness and reliability. The Inception-V3 model achieved impressive accuracies of 97.48% on the Dataset1, 98.23% on Dataset2, and 97.85% on Dataset3, demonstrating its effectiveness in distinguishing between genuine and forged signatures