Analyzing Artifacts in Discriminative and Generative Models Richard - PowerPoint PPT Presentation

Dataset of CNN-generated fakes Generative Adversarial Networks Perceptual loss Low-level vision Deepfakes synthetic real ProGAN Pr Styl yleGAN BigGAN Bi AN Cyc ycleGAN St StarGAN GauGA Ga GAN Casc scad aded ed re refine IM IMLE Se Seeing g in th the d dark Su Supe perres Facesw swap Karras 2018 Karras 2019 Brock 2019 Zhu 2017 Choi 2018 Park 2019 Chen 2017 Li 2019 Chen 2018 Dai 2019 Rossler 2019 Trai Tr ain Test st Many di differ eren ences ces (architecture, dataset, objective) …but underlying co commonal aliti ties es may enable generalization

CNN-generated images are surprisingly easy to spot…for now Sheng-Yu Wang Oliver Wang Richard Zhang Andrew Owens Alexei A. Efros In CVPR CVPR , 2020 (oral).

Training on ProGAN ProGAN detector ProGAN-generated !(#) % & Real vs. fake 720K real images, 20 categories from LSUN

Testing across architectures ProGAN detector Synthesized images from ot other her CNNs … & Real vs. fake … Real images

Average Generalizes above chance, but performance inconsistent Precision Perfect 100 100 100 100 99 99 98 98 97 96 96 95 Training and testing on ProGAN is trivial 94 94 93 90 88 84 72 67 66 64 No augmentation Chance Blur+JPEG aug (at training) ProGAN IMLE StyleGAN CRN GauGAN CycleGAN StarGAN Seeing dark BigGAN Deep fake Super-res.

Augmentation is not always appropriate Average Precision Perfect 100 100 100 100 99 99 98 98 97 96 96 95 94 94 93 90 88 84 72 67 66 64 No augmentation Chance Blur+JPEG aug (at training) ProGAN IMLE StyleGAN CRN GauGAN CycleGAN StarGAN Seeing dark BigGAN Deep fake Super-res.

Aggressive augmentation adds surprising generalization Average Precision Perfect 100 100 100 100 99 99 98 98 97 96 96 95 94 94 93 90 88 84 72 67 66 64 No augmentation Chance Blur+JPEG aug (at training) ProGAN IMLE StyleGAN CRN GauGAN CycleGAN StarGAN Seeing dark BigGAN Deep fake Super-res.

Average Precision Blurring at test-time Perfect 100 100 100 100 99 99 98 98 97 96 96 95 94 94 93 Indicates exploitable, generalizable 90 88 wer frequency bands artifacts at lo lowe AP 84 72 ' 67 66 64 No augmentation Chance Blur+JPEG aug (at training) ProGAN IMLE StyleGAN CRN GauGAN CycleGAN StarGAN Seeing dark BigGAN Deep fake Super-res.

Discussion • Suggests CNN-generated images have common artifacts • Artifacts can be detected by a simple classifier! • StyleGAN2 (released af after er our submission): 100% AP on FFHQ • Swapping Autoencoder (Park et al.): 95% AP on FFHQ • No Note : AP is computed on a collection of images; a real/fake decision on a per-image basis is more difficult • Situation may not persist • GANs train with a discriminator • Future architecture changes • “Shallow” fakes, e.g., Photoshop 75

Media manipulation example https://gizmodo.com/russian-state-tv-photoshops-an-awkward-smile-on-kim-jon-1826529277 June 2018

https://www.youtube.com/watch?v=5Qqv_C6iVvQ

Original

#1 modification

#2 modification

#3 modification

#4 modification

Photoshop Undo-er Photoshopped Original Or Origin/PS’ PS’d & Fl Flow field Warp Recovered original

Photoshop Undo-er “G “Ground und trut uth” h” PWC-Net fl flow ow fi fiel eld Original Loss & Flow field Fl Loss Warp Recovered original

Manipulated

Flow prediction

Suggested “undo” Prediction Suggested “undo”

Original

Manipulated vs. Original

Suggested “undo” Prediction Manipulated Undo vs. Original

Manipulated

Suggested “undo” Prediction Manipulated Flow prediction

Suggested “undo” Prediction Suggested “undo”

Original

Manipulated vs. Original

Suggested “undo” Prediction Manipulated Undo vs. Original

Reversal evaluation Held-out artist generated data Senses of generalization • Heldout artist data 42.7 PSNR (dB) 31.2 29.0 Before Prediction GT Flow

Reversal evaluation Facebook post-processing Senses of generalization +0.61 dB • Heldout artist data • Post-processing • Different warp, image domains +0.15 dB No aug With Aug Data augmentation important (again)

Snapchat warps Original Photo

Snapchat warps Manipulated Photo

Snapchat warps Suggested “undo” Prediction Manipulated Flow Prediction

Snapchat warps Suggested “Undo”

Snapchat warps Some generalization across warp methods Original Photo

Different image domain