A Robust Coverless Image Steganography Algorithm Based on Image Retrieval with SURF Features

Table 5

Robustness comparison with LDA_DCT [18], DenseNet_DWT [19], and DCMH-CNN [32] in ImageNet dataset.

Processing	Size	LDA_DCT (%)	DenseNet_DWT (%)	DCMH-CNN (%)	Proposed (%)
JPEG	Q (10)	91.4	98.0	59.4	98.8
	Q (50)	98.4	98.0	94.1	100.0
	Q (90)	99.6	100.0	100.0	100.0
Gauss-N	σ (0.001)	94.1	98.8	99.6	100.0
	σ (0.005)	94.1	98.0	93.0	100.0
S&P-N	σ (0.001)	98.4	100.0	100.0	100.0
	σ (0.005)	94.1	98.8	99.6	100.0
Speckle-N	σ (0.01)	99.6	99.2	99.6	99.6
	σ (0.05)	89.8	91.0	73.8	98.4
Gauss-F	(33)	99.2	100.0	100.0	100.0
Mean-F	(33)	98.8	100.0	97.7	100.0
Median-F	(33)	94.5	95.7	98.8	100.0
Centered-C	20%	66.8	20.3	98.4	94.5
	50%	11.3	3.5	66.0	48.0
Edge-C	10%	18.8	55.5	97.3	100.0
	20%	6.3	28.5	90.6	92.6
Rotation	10°	66.4	36.7	100.0	98.8
	30°	8.2	5.9	93.8	94.1
	50°	5.1	2.3	70.3	95.7
Translation	(80, 50)	21.1	36.7	98.4	98.8
	(160, 100)	8.6	9.0	88.3	94.1
	(320, 200)	6.3	5.9	70.3	91.8
Scaling	0.5	97.7	92.2	87.5	98.4
	0.75	96.9	94.9	100.0	100.0
	1.5	99.2	98.8	100.0	100.0
	3	98.8	100.0	100.0	100.0
C-H-E		73.4	71.1	96.5	75.0
Gamma-C	0.8	89.8	94.5	100.0	100.0
Average		68.8	69.0	91.2	95.7

The bold data in the table represent the highest accuracy among the compared algorithms. The italicized data represent the second-highest accuracy.