Offline signature verification remains the most commonly employed authentication modality and enjoys global acceptance. From the view point of computerized verification, concluding the authenticity of a signature offers a challenging problem for the pattern classification community. A major proportion of computerized solutions treat signature verification as a two-class classification problem where both genuine and forged signatures are employed for training purposes. For most of the real world scenarios however, only genuine signatures of individuals are available. This paper presents a signature verification technique that relies only on genuine signature samples. More precisely, we employ convolutional neural networks for learning effective feature representations and a one-class support vector machine that learns the genuine signature class for each individual. Experiments are carried out in a writer-dependent as well as writer-independent mode and low error rates are reported by only employing genuine signatures in the training sets.