<p>Handwritten signatures are considered unique for a person’s identity and hence are used as a prevalent method of identification in banking, business, financial settings, legal documents, etc. No two signatures of a person are exactly the same; slight variations occur each time one signs. Taking advantage of the variations present, forgers may forge an individual’s signature. This leads to the requirement of developing a system which learns internal patterns in a signature and can accurately classify between forged and original signatures. Various machine learning and deep learning methods have been explored, but machine learning methods are dependent on manual feature extraction, which makes it difficult to generalise. At the same time, deep learning methods can perform the same task automatically, but require vast amounts of training data, which is small in this case. This manuscript proposes a robust method, SignaLearn, for signature verification using a pre-trained VGG19 network for feature extraction and machine learning classifiers for accurate classification. By employing this strategy, we achieve the best of both worlds, enabling us to train networks on limited data while verifying signatures using features extracted from VGG19. The proposed method is validated on three publicly available benchmark datasets: BHSig260 (Bengali script), CEDAR, and UTSig (Persian signatures). The proposed method ensures that the classifiers are well adapted to the complex feature space of signature data. The proposed model demonstrates strong generalisation across datasets, achieving an accuracy of 83.703% on BHSig260 (Bengali), 100% on CEDAR, and 87.983% on UTSig. Subject-wise accuracies were also evaluated for three of the datasets, achieving, on average, perfect accuracy of 100%. These results indicate that the combination of transfer learning with VGG19 and optimised machine learning classifiers provides a powerful solution for offline signature verification, even in scenarios with limited annotated data.</p>