Deep Learning-Based Forgery Attack on Document Images

With the ongoing popularization of online services, digital document images have been used in various applications. Meanwhile, there emerged some deep learning-based text editing algorithms which alter textual information an image . In this work, we present a forgery algorithm to edit practical images. To achieve goal, limitations existing towards complicated characters and complex background are addressed by set network design strategies. First, unnecessary confusion supervision data is avoided disentangling source Second, capture structure components, skeleton provided as auxiliary continuity texture considered explicitly loss function. Third, traces induced operation mitigated post-processing operations consider distortions from print-and-scan channel. Quantitative comparisons proposed method exiting approach shown advantages our reducing about 2/3 reconstruction error measured MSE, improving quality PSNR SSIM 4 dB 0.21, respectively. Qualitative experiments confirmed that results visually better than approach. More importantly, demonstrated performance under scenario where attacker able identity using only one sample target domain. The forged-and-recaptured samples created attack recapturing successfully fooled authentication systems.