Overview of Image Inpainting and Forensic Technology

Table 6

Comparison of GAN-based forensics.

Category	Source	Detection type	Methods	Performance
Classification	[152]	Deep network-generated images	Residual domain of chrominance components	100% (ACC)
	[153]	Cycle-GAN Star-GAN	Co-occurrence matrices	93.42 (accuracy) 99.49 (accuracy)
	[154]	GAN-generated images	INHNet Saturation feature	0.56 (AUC) 0.7 (AUC)
	[155]	Cycle-GAN, ProGAN, Star-GAN	GAN fingerprints	0.999 (AUC)
	[156]	GAN-generated images	GAN fingerprints	99.93% (accuracy)
	[157]	11 GAN models	Single ProGAN	93.0 (mAP)
	[158]	PGGAN	Lap-CNN	96.3%
	[159]	Cycle-GAN	Eight methods	>83.58%
	[160]	GAN-generated human face images	CNN with self-attention mechanism	99.3 (accuracy)
	[161]	Five GAN models	EM; fingerprint	99.65 (accuracy)
	[162]	PGGAN	Shallow CNN architecture	99.99% (AUC)
	[163]	PGGAN	Xception	99.99% (accuracy)
	[164]	PGGAN, WGAN, Style-GAN, LSGAN, DCGAN	Global and local feature, ArcFace loss, CNN	99.99% (accuracy)
	[165]	BigGAN, Style-GAN2, PGGAN	R, G, and B components, DWT, SVM	98.45% (accuracy)
Localization	[166]	Criminisi, SN-PatchGAN	Multitask deep learning network based on mask R-CNN	97.8 (mAP)
	[167]	GAN-based face manipulation	Gray-scale fakeness prediction map; encoder-decoder architecture with attention mechanism	99.95 (ACC)