Audio Steganalysis with CNN

@ author: Wang Yuntao

Necessary Package

tensorflow-gpu==1.3, numpy, matplotlib

CNN Architecture (To be perfected)

Steganographic Algorithm

HCM(Huffman Code Mapping), EECS(Equal Length Entropy Codes Substitution)

Results (To be perfected)

• HCM-Gao steganographic algorithm

Bitrate	RER	Ours	Ren[1]	Jin[2]
128 kbps	0.1	70.72	50.13	50.11
	0.3	75.18	51.41	52.34
	0.5	78.53	53.75	52.79
320 kbps	0.1	73.83	50.77	50.85
	0.3	77.27	52.18	52.63
	0.5	80.71	55.69	54.42

• HCM-Yan steganographic algorithm

Bitrate	RER	Ours	Ren[1]	Jin[2]
128 kbps	0.1	75.92	50.94	51.25
	0.3	81.39	56.88	60.31
	0.5	85.88	76.56	52.79
320 kbps	0.1	79.35	64.38	50.85
	0.3	83.09	66.56	52.63
	0.5	90.21	77.81	54.42

• EECS steganographic algorithm

Bitrate	RER	Ours	Ren[1]	Jin[2]
128 kbps	2	92.39	59.42	56.25
	4	83.17	51.37	51.81
	6	75.12	50.96	50.73
	7	63.78	50.56	50.35
	8	51.37	50.10	50.07
320 kbps	2	95.35	71.38	69.69
	4	93.09	52.81	52.56
	6	85.33	51.24	51.13
	7	64.46	50.10	50.07
	8	51.42	50.00	50.00

file description

ID	File	Function
1	audio_preprocess.py	include some pre-process methods for audio
2	text_preprocess.py	include some pre-process methods for test
3	image_preocess.py	include some pre-process methods for image
4	file_preprocess.py	get the name, size and type of the file
5	pre_process.py	some pre-processing method such as truncation, down_sampling
6	classifier.py	machine learning classifiers such as SVM, KNN, and model selection, ROC plot, etc.
7	config.py	command parser and some package management
8	filters.py	some filters used for pre-processing such as kv kernel or other rich model
9	main.py	the main program
10	manager.py	GPU management (free GPU selection automatically)
11	layer.py	basic unit in CNN such as conv layer, pooling layer, BN layer and so on
12	network.py	various networks including VGG19, LeNet and ourselves' network
13	train.py	the train of the network
14	test.py	the test of the network
15	utils.py	some useful tools such as minibatch, get_model_info, get_weights, get_biases and so on
16	figure.py	visualization analysis
17	TODO	to do list
18	model	model files Folder
19	label.txt	label file if batch test

Run

• install python3.x and add the path into environment variable
• GPU run enviroment configure if train the network (optional)
• pip install tensorflow==1.3 numpy scikit-image, pydub (depend on FFmpeg, optional)
• run the code as the example as follows

Command Parser

Command: python3 main.py --argument 1 --argument 2 ... --argument N

Namespace(batch_size=128, batch_size_train=128, batch_size_valid=128, bitrate=128, block=2, carrier='audio', cover_train_dir=None, cover_valid_dir=None, data_dir=None, decay_method='exponential', decay_rate=0.9, decay_step=5000, direction=0, downsampling=False, end_index_train=-1, end_index_valid=-1, epoch=500, height=512, is_abs=False, is_diff=False, is_regulation=True, is_trunc=False, keep_checkpoint_every_n_hours=0.5, learning_rate=0.001, log_dir=None, logs_path='/home/zhanghong/code/CatKing/steganalysis_CNN/logs', max_to_keep=3, mode='test', model_dir='/home/zhanghong/code/CatKing/steganalysis_CNN/models/stegshi', model_file_name='audio_steganalysis', model_file_path=None, models_path='/home/zhanghong/code/CatKing/steganalysis_CNN/models', network='stegshi', order=2, relative_payload='2', seed=1, staircase=False, start_index_train=0, start_index_valid=0, stego_method='EECS', stego_train_dir=None, stego_valid_dir=None, submode='one', test_file_path='/home/zhanghong/data/image/val/512_stego/result_7518.pgm', test_files_dir=None, threshold=15, width=512)

Example:

--train

simple 1: sudo python3 main.py --mode train --data\_dir /home/"home_name"/data --height 200 --width 380

simple 2: sudo python3.5 main.py --mode train --carrier image --height 512 --width 512 --network stegshi --batch_size 64 --end_index_train 6000 --end_index_valid 1500 --cover_train_dir /home/zhanghong/data/image/train/512_cover --cover_valid_dir /home/zhanghong/data/image/val/512_cover/ --stego_train_dir /home/zhanghong/data/image/train/512_stego/ --stego_valid_dir /home/zhanghong/data/image/val/512_stego/ --logs_path /home/zhanghong/code/CatKing/steganalysis_CNN/logs/stegshi --models_path /home/zhanghong/code/CatKing/steganalysis_CNN/models/stegshi

--test

sample 1: sudo python3.5 main.py --mode test --submode one --network stegshi --model_dir /home/zhanghong/code/CatKing/steganalysis_CNN/models/stegshi --file_path /home/zhanghong/data/image/test/12138.pgm

sample 2: sudo python3.5 main.py --mode test --submode one --network stegshi --height 512 --width 512 --model_dir /home/zhanghong/code/CatKing/steganalysis_CNN/models/stegshi --test_file_path /home/zhanghong/data/image/val/512_cover/7501.pgm

sample 3: python3 main.py --mode test --submode one --network stegshi --height 512 --width 512 --model_file_path stegshi/audio_steganalysis-5797 --test_file_path TEST/7501.pgm

sample 4: python3 main.py --mode test --submode batch --network stegshi --height 512 --width 512 --model_file_path stegshi/audio_steganalysis-5797 --test_files_dir TEST/

sample 5: python3 main.py --mode test --submode batch --network stegshi --height 512 --width 512 --model_file_path stegshi/audio_steganalysis-5797 --test_files_dir TEST/ --label_file_path label.txt


Note: remove "sudo" if run the code in windows system

arguments:
Mode: mode, test, data_dir
Data_info: bitrate, relative_payload, stego_method, model_dir, log_dir
Hyper_paramsters: batch_size, batch_size_train, batch_size_valid, epoch, learning_rate, gpu, seed, decay_step, decay_rate, staircase
Path: cover_train_dir, cover_valid_dir, stego_train_dir, stego_valid_dir, models_path, logs_path
Network: network

illustration:
mode: train | test
test: one | batch
data_dir: data dir
bitrate: 128 | 192 | 256 | 320
stego_method: EECS | HCM-Gao | HCM-Yan
relative_payload:

2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10, if the steganography algorithm is the EECS algorithm
1 | 3 | 5, otherwise

The introdction of each network

• network for audio steganalysis

    network1  : The proposed network (最终选定的网络)
    network1_1: Remove the BN layer (去掉BN层)
    network1_2: Average pooling layer is used for subsampling (将所有的降采样方式改为平均池化方式)
    network1_3: Convolutional layer with stride 2 is used for subsampling (将所有的降采样方式改为卷积池化方式)
    network1_4: Replace the convolutional kernel with 5x5 kernel (将卷积核尺寸由3 x 3改为5 x 5)
    network1_5: ReLu is used as the activation function (将激活函数由Tanh改为ReLu)
    network1_6: Leaky-ReLu is used as the activation function (将激活函数由tanh改为Leaky-ReLu)
    network1_7: Deepen the network to block convolution layers (加深网络)
    network1_8: Design a network to steganalyze audios of arbitrary size (解决可变尺寸输入数据的训练问题)
    
    Note: HPF and ABS is applied at the pre-processing
  

• network for image steganalysis

    stegshi   : Xu-Net
  

• The method of pre-processing There are positive and negative values in QMDCT coefficients matrix. The values in interval [-15, 15] is modified. The ratio of values in [-15, 15] is more than 99%, as the figure shown.

  • Abs
  • Truncation
  • Down-sampling

• Data Folder Name
  

  cover

    128 | 192 | 256 | 320
  

  stego

    EECS
        128_W_2_H_7_ER_10, 128_W_3_H_7_ER_10, ..., 128_W_10_H_7_ER_10
        192_W_2_H_7_ER_10, 192_W_3_H_7_ER_10, ..., 192_W_10_H_7_ER_10
        256_W_2_H_7_ER_10, 256_W_3_H_7_ER_10, ..., 256_W_10_H_7_ER_10
        320_W_2_H_7_ER_10, 320_W_3_H_7_ER_10, ..., 320_W_10_H_7_ER_10
    W: The width of the parity-check matrix, W = 1 / α, and α is the relative payload
    H: the height of the parity-check matrix
    ER: The number of fremes used for embedding = The number of whole frames * ER
    
    HCM-Gao
        128_01, 128_03, 128_05, 128_10
        192_01, 192_03, 192_05, 192_10
        256_01, 256_03, 256_05, 256_10
        320_01, 320_03, 320_05, 320_10
    01, 03, 05, 10 is the ER as shown above
    
    HCM-Yan
        128_01, 128_03, 128_05, 128_10
        192_01, 192_03, 192_05, 192_10
        256_01, 256_03, 256_05, 256_10
        320_01, 320_03, 320_05, 320_10
  

Reference

[1] Yanzhen Ren, Qiaochu Xiong, and Lina Wang. 2017. A Steganalysis Scheme for AAC Audio Based on MDCT Difference Between Intra and Inter Frame. In Digital Forensics and Watermarking - 16th International Workshop, IWDW 2017, Magdeburg, Germany, August 23-25, 2017, Proceedings. 217–231.
[2] Chao Jin, Rangding Wang, and Diqun Yan. 2017. Steganalysis of MP3Stego with low embedding-rate using Markov feature. Multimedia Tools and Applications 76, 5 (2017), 6143–6158.