Deep 3D mesh watermarking with self-adaptive robustness

Cybersecurity

Table 1 Quantitative comparisons with baseline methods on Hand and Asiadragon dataset

	Method	HD	MRMS	\(l_{cur}\)	Accuracy (%)	Emb time (s)	Ext time (s)
Asiadragon	Bin (Cho et al. 2007)	0.0027	0	0	74.47	0.406	0.015
	L-M (Bors and Luo 2012)	0.0027	0	0	76.62	2547	0.015
	Laplacian (Cayre et al. 2003)	0.0319	0.012	0.030	94.34	0.725	0.566
	MAPS (Liu et al. 2017)	0.0002	0	0	75.41	0.004	0.003
	SCKM (Lee et al. 2021)	0	0	0	85.39	0.012	0.012
	Proposed	0.0498	0.014	0.001	95.22	0.032	0.016
Hand	Bin (Cho et al. 2007)	0.0032	0	0	71.84	0.327	0.013
	L-M (Bors and Luo 2012)	0.0063	0	0	70.09	1040	0.013
	Laplacian (Cayre et al. 2003)	0.0208	0.006	0.028	91.22	0.523	0.347
	MAPS (Liu et al. 2017)	0.0001	0	0	74.25	0.010	0.007
	SCKM (Lee et al. 2021)	0	0	0	86.34	0.011	0.011
	Proposed	0.1131	0.018	0.003	92.06	0.022	0.012

Distance between original meshes and watermarked meshes (second to fourth column), bit accuracy under attacks from the attack layers (%, fifth column) and running time (sixth to seventh column). Running time consists of the watermark embedding time and watermark extracting time for one 3D mesh. For all indicators the lower the better except accuracy
The bold represents the ones with best preformance