<p>Offline handwritten signature verification systems require several authentic samples to achieve high performance. However, when developing such systems, customers provide few samples of their authentic signatures. For this reason, researchers employed data augmentation techniques in order to increase the number of genuine signatures. As both real and synthetic signatures undergo a feature generation process, applying the data augmentation in the feature space reduces the computation cost and could be more effective than duplicating signature images. In this work, the proposed synthetic feature generator is based on a 1D convolutional GAN (Generative Adversarial Network). Precisely, the 1D-GAN is designed to generate synthetic features from authentic features derived from various signature descriptors. Three different descriptors are employed to ensure diversity when evaluating the 1D-GAN performance. The first descriptor is based on the SigNet which generates deep features while the two others are handcrafted descriptors based on the Local Directional Patterns and the Histogram of Templates. The verification step is achieved by a SVM classifier. Experiments conducted on CEDAR and MCYT-75 datasets, reveal that synthetic features allow a significant improvement of the verification scores when the system is developed by using one real signature. In this respect, the proposed system associated with SigNet features, emphasizes a gain of 20% and 11% in the AER for both CEDAR and MCYT-75 datasets, respectively.</p>