Generative Adversarial Network for Identifying Authors of Traced Images

Art forgery is one of the important issues in art, but the general approaches using well-known image datasets are likely to fail to identify real forgeries. This is because the distributions of the well-known image datasets are very different from that of a dataset that includes the pictures created by professional counterfeiters. However, there are few datasets specialized for art forgery detection. In this research, we constructed a dataset for author identification of drawings. To create the drawings, we traced a few original sketches on top of semi-transparent papers. The traced images look more similar to the originals than typical forgery artwork images do, and it implies that it is more difficult to identify the authors of the traced images. We also suggest a new generative adversarial network model for author identification. We approach the problem by maximizing mutual information between an author and his traced artwork image. We compared our results to those obtained by the baseline convolutional neural networks. The experimental results showed that the suggested approach produced more convincing results than other approaches.