F3SNet: A Four-Step Strategy for QIM Steganalysis of Compressed Speech Based on Hierarchical Attention Network

Table 2

Detection results for different length samples (embedding rate: and ).

Embedding rate (%)	Sample length (s)	ACC (%)	FNR (%)	FPR (%)
20	0.1	70.12	47.328	12.266
	0.2	74.71	37.451	13.417
	0.4	76.46	32.393	14.608
	0.6	80.18	25.911	15.306
	0.8	81.59	21.816	14.694
	1.0	83.45	16.488	16.618
	2.0	90.58	11.868	6.961
	4.0	94.63	4.485	6.3
	5.0	95.46	5.769	3.274
40	0.1	83.98	25.882	6.225
	0.2	87.11	17.083	8.549
	0.4	90.53	11.874	7.094
	0.6	92.48	8.847	6.238
	0.8	95.07	5.058	4.804
	1.0	94.63	4.762	5.962
	2.0	98.14	2.649	1.069
	4.0	99.66	0.294	0.390
	5.0	99.90	0	0.194