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

Table 1

Experiment with different types of network models.

Number	Network model	Hyperparameters
Model #1	F3SNet	The dimension of embedding layer = 100, the number of word LSTM hidden unit = 100, the number of sentence LSTM hidden unit = 50, dropout = 0.5, dropout_recurrent = 0.5, batch size = 128, and epoch = 50
Model #2	Embedding + LSTM + Self_Attention + Dense	The dimension of embedding layer = 100, the number of LSTM hidden unit = 100, dropout = 0.5, dropout_recurrent = 0.5, batch size = 128, and epoch = 50
Model #3	Embedding + Self_Attention + Self_Attention + Dense	The dimension of embedding layer = 100, dropout = 0.5, batchsize = 128, epoch = 50.
Model #4	LSTM + Self_Attention + BiLSTM + Self_Attention + Dense	The number of word LSTM hidden unit = 100, the number of sentence LSTM hidden unit = 100, dropout = 0.5, dropout_recurrent = 0.5, batch size = 128, and epoch = 50
Model #5	Embedding + Multi-head Attention + Dense ([22])	The dimension of embedding layer = 100, heads = 8, head_size = 32, dropout = 0.5, batchsize = 128, epoch = 50.
Model #6	LSTM + LSTM + Dense ([20])	The number of the first LSTM hidden unit = 50, the number of the second LSTM hidden unit = 50, batch size = 128, and epoch = 50