Variant length , Self-extracted audio watermark for verification using LWT and random selections

In the last decade with the expansion of cyber multimedia activities, concepts like authentication, identification and verification became a must. Audio is one of the challenging media in cyber security for its complicated nature. Watermarking rises as an important methods used in securing audio files and other media. In this research a new method is used for extracting the signal features from random positions in the original audio signals by some signal calculations in time domain and hide them within the same audio in other positions after transforming the samples in these positions using lift wavelet transform, all positions were chosen depending on random walk method and a secret key. The extracted features will be compared with the hidden features (watermark) for verification. The proposed method was tested against compression (mp3) and noise addition (White Gaussian noise). Many types of performance measurements like peak signal to noise ratio, bit error rat, mean square error and others were used to measure the efficiency of the proposed method.

should be easily and correctly extracted by the authorized receiver and in the same time should be difficult to be detected by attackers.Any type of data could be used as a watermark (audio, image,…) Since the watermark is hidden in the original file, it could be used for many purposes like authentication, verification, identification and recognition.[1] A lot of watermarking methods use the original audio data (samples in time domain) to hide the watermark which cause the watermark to be fragile and easy attacked and detected.Therefore, all new researches went toward hiding watermark in a transformed representation of the audio data.Transformers like DCT, FFT, SFFT, WLT, DWT and others were used.By using transformers, watermark is distributed in a large spectrum (SS) which makes the watermark immune against many types of attacks.The wavelet methods calculate differences and averages of audio signal, breaking down the signal into spectrum.Wavelet transform gave two groups of values: a group of averages and a group of differences (the second group is called wavelet coefficients).The size of the averages group and the coefficients group is half the size of the original data.For example, if the time series contains 512 samples, the first and second group will be of size 256 each.Most transformers use data of power 2, the spectrum results from a wavelet process reflect the modification in the time series with different resolutions.The first coefficient reflects the largest change in frequency.All later coefficients reflect modifications at lower frequencies.[2] In lifting transform, the wavelet finds the difference between the predicted values and an original values.
The LWT uses a different calculation which is close to Haar as in eq.The LWT is a wavelet algorithms which has excellent reconstruction and the best multi-scale resolution.[3] In ths paper, we focus on audio signals such as music records.Audio watermarking is the process of hiding special data (such as features, names, images, etc.) into the original audio without affecting it.[4] A successful and practical watermark scheme should satisfy some important issues such as robustness, imperceptibility, security and great embedding capacity.Blind watermarking methods, which are able to extract the watermark without having the original audio signal are more practical and desired in cyber security.[5] The proposed method combines many concepts like LWT, random walk, variant length watermark, and blind extraction method, to present a new powerful, practical, light, and immune watermarking strategy.

The Proposed method
The proposed method is composed of many parts: (position selection, feature extraction, watermark hiding) for sender side, and (position selection, feature extraction, watermark extraction and comparison) for receiver side.Each part will be explained in details.

2-1-1 Position Selection:
This is the first step in the process of watermarking.Locations to extract the features shouldn't be fixed or pre-determined, they must be variant with each new audio file to prevent predicting attack.The variations come not only from the random walk(RW) function, but also from using Each frame will have its own feature matrix that differs from all other frames.The whole feature matrices will represent the characteristics of the audio file which could be used for verification of the audio or even for identification if the audio file is a personal voice.

2-1-3-Feature Hiding:
After extracting the watermark from some random selected blocks(frames), they will be hidden in other random selected blocks.Hiding process is more complicated than the extraction process, since the hidden data should not affect the sound and should not be perceptible or predictable by attackers.3), a watermarked sound file was created and ready to be sent.

2-2 Receiver:
Since the proposed watermarking algorithm is blind, the receiver doesn't have the original audio file; only the watermarked file is received; he has to apply the same algorithms(1, 2, 3), using the same secret key.
By applying algorithm(1) and ( 2), the receiver extract the features and is ready to continue with the watermark extraction.

Implementation
The proposed method was applied on different audio files.The following implementation is executed on one audio file (type: mono, frequency: 11025 Hz, duration: 27sec).Five watermarks were hidden within that file; all watermarks were extracted from the same audio file from different frames; each watermark represents the features of one frame.In each figure (2:6), the third plot represents the overlapping between the original file and the watermarked file each with different color (red for original and green for watermarked).It is very clear that the differences between them were tiny so they only sometimes appeared like red dots on the green signal.
The watermarks data which extracted from the frames of each sound file (here 5 frames), were all differ from each other, that's why the proposed method could be used not only as a protection watermarking method but also as a method for verification, identification and even recognition for human voices.Figures (8:16) show the watermarks extracted from each audio file, the differences were clarified by plotting the watermark data.IV.
The proposed method applied on audio files for cyber transmission, which means it has to be light and powerful.Applying LWT only on the frames where the watermark was hidden and not on all audio file makes the process fast and efficient. V.
The positions used for feature extraction and hiding changed with the secrete key which was used in random walk function to produce positions; therefore, the watermark could be extracted only by knowing the key otherwise, the watermark is secured.

VI.
Using variant (length and value) watermark gave the algorithm more power and resistant against attack since it is difficult to predict the positions and the length of each watermark.
VII.The watermark is found from the audio itself which makes the algorithm really blind.The attacker has no pervious knowledge about the watermark.Also, the watermark describes the characteristics of the audio signal which is unique and could be used for verification, identification and recognition.
VIII.More than one watermark is hidden within each audio file (No. of WM), and that helps when part off the audio file is corrupted, the rest of the file may be checked.All tested audio files have been attacked by compression (mp3) and noise addition (White Gaussian noise), in all cases the watermarks were successfully extracted.

X.
The SNR values for the tested audio files were all greater than 70% depending on the size of the watermark which itself depends on the nature of the signal.

Conclusion
Cyber security is one of the most important concepts these days which made all the effort towards securing transmission and data.Audio files are most challenging media for its complicated and fragile nature.To secure audio files, a light and powerful method is needed.

Algorithm ( 3 2 -
Convert each watermark to one dimensional binary vector.(Note: each WM has different length).Vol: 13 No:2 , April 2017 DOI: http://dx.doi.org/10.24237/djps.1302.241CP-ISSN: 2222-8373 E-ISSN: 2518-9255 3-For each hiding frame (of size = WM), Convert using lift wavelet transform (type Haar); 4-Take only the coefficient vector of LWT and convert to binary.5-Replace the least significant bit LSB with one bit of the watermark binary vector.6-Apply inverse transform (ILWT); 7-Return the sound frame to its original place in the sound file; 8-End By the end of algorithm(

Algorithm( 4 2 -
For all features extracted by algorithm(2) on watermarked received audio file, convert to one dimensional binary vector.3-For each hiding frame (of size = WM), convert using lift wavelet transform.Take only the coefficient vector of LWT; 4-Convert these coefficients to binary; 5-Compare the LSB of each binary coefficient with the binary feature vector; if they are all equal, then the audio file is verified.6-End Vol: 13 No:2 , April 2017 DOI: http://dx.doi.org/10.24237/djps.1302.241CP-ISSN: 2222-8373 E-ISSN: 2518-9255

P
Figure 12: the plot representation of watermark data(Whitney.wav)