One of the major limitations which obstruct the development of an efficient offline signature verification system is the lack of genuine signatures required to perform a reliable training stage. To deal with this issue, data augmentation techniques are employed to add synthetic genuine signatures. Presently, the signature verification system is developed by applying an algorithm for synthetic features generation. We evaluate the effectiveness of this feature generator with two different signature descriptors. The first descriptor is the Local Directional Strength Patterns while the second descriptor is based on the SigNet deep features. The verification phase is carried out with an SVM classifier. Experiments are performed on CEDAR and MCYT-75 datasets according to the writer-dependent verification approach. The inclusion of synthetic features considerably enhances the system performance, yielding a gain of 17% when the verification system employs only one real signature. Compared to handcrafted features, the SigNet provide more effective synthetic features with both datasets.