[Retracted] Deep Image Watermarking to JPEG Compression Based on Mixed-Frequency Channel Attention

Table 2

Results of robustness against other distortions. We add an additive diffusion block into the message processor for improving the shortcoming that our original structure is not robust enough against crop, crop out, and drop out and make a comparison with [26–28] trained by a combined noise layer. The strength factor is adjusted for comparison under .

Noise	Identity	Crop out	Drop out	Crop	GF	JPEG
HiDDeN [26]	0%	6%	7%	12%	4%	37%
TSDL [27]	0%	2.7%	2.6%	11%	1.4%	23.8%
MBRS [28]	0%	0.0027%	0.0087%	4.15%	0.011%	4.48%
Ours	0%	0.0013%	0.0080%	3.24%	0.293%	2.61%