2020 IEEE 5th International Conference on Computing Communication and Automation (ICCCA)

Galgotias University, Greater Noida, UP, India. Oct 30-31, 2020

A Robust MP3 Audio Steganography with

Improved Capacity
Hemalatha S Ramathmika
Department of Computer Science and Engineering Department of Computer Science and Engineering
Manipal Institute of Technology, Manipal Academy of Higher Manipal Institute of Technology, Manipal Academy of Higher
Education, Manipal Education, Manipal
Karnataka, India-576104 Karnataka, India-576104
[email protected] [email protected]

Abstract— Data transmission on the internet should be are not safe for steganography, if the stego signal has to
protected from severe threats. An effective and efficient withstand compression. Thus, the only available portion is
method to achieve this is Steganography. Steganography is a the header. The possible fields available for steganography in
type of secret communication in which the secret information is the header are padding byte, unused header bits such as
hidden in a carrier such as an image, audio, or a video. Audio
files are very often used by the steganographers as the carriers
private bit, copyright bit, original bit, mode extension bits
of secret information because of the large amount of space and emphasis bits. Mode extension bits can be used if the
available for information hiding. This paper proposes a robust channel mode is not set to the stereo mode. Similarly,
steganography technique applied to MP3 audio with improved overwriting emphasis bits produces distortion on, some audio
capacity. The proposed method is robust against compression, players. Another field that can be used to hide information is
cropping, filtering, resampling and normalization. The ancillary data. Thus, a meagre number of bits per frame is
performance of the technique is measured using the metrics - available to be used for steganography. The challenge in
Signal to Noise Ratio (SNR) and Squared Pearson Correlation MP3 audio steganography is to get sufficient cover audio bits
Coefficient (SPCC). that can be modified to hide secret information and to extract
Keywords— Audio steganogrphy, robust, MP3, capacity,
the secret bits without loss.
information hiding, signal to noise ratio, Squared Pearson
II. LITERATURE REVIEW
Correlation Coefficient
Most of the data hiding methods for MP3 audio that have
I. INTRODUCTION been proposed in the literature have low embedding capacity.
In audio steganography, the carrier which is called the For MP3 audio, MP3Stego is the most widely used tool
cover object is an audio file. The secret information may be a among a few stego tools viz., MP3Stego, UnderMP3Cover,
text file, an image or an audio file. Security and capacity are and MP3Stegz. It has been proved that MP3Stego detection
the two important characteristics of steganography. Security is possible [3]. In [3], the authors proposed an MP3
is the ability of the steganography technique to be undetected steganography method by exploiting the rule of window
by eavesdroppers. Metrics such as Signal to Noise Ratio switching during encoding, which exhibits a very low
(SNR) and Squared Pearson Correlation Coefficient (SPCC) capacity. The complexity involved in the MP3 audio
are used for measuring the security. Capacity is the amount steganography is by its format [1], [2]. In MP3 files, the
of secret information hidden in the cover object. It is information hiding process may be carried out before, during
measured as a percentage of the cover object. However, MP3 or after compression. Pre-compression hiding is risky as it
audio steganography is challenging because of its may lead to loss of hidden data. Various compression steps
compressed form. The secret information should be retained may be used for data hiding if complexity and the capacity
under compression. Usually, the audio files are stored in a are not an issue. MP3 files can be watermarked using
compressed form, and the most commonly used form is watermarking algorithms that are robust to compression [4],
MP3. When the audio frames are used for steganography, [5]. The redundant bits in the frames are manipulated to
appropriate frames should be selected to hide the secret embed the watermark during the encoding process. Barnali
information, such that the secret information remains Gupta Banik et al. proposed an audio steganography
unharmed during the compression. [1], [2]. technique that uses both Discrete Wavelet Transform (DWT)
An MP3 file consists of several frames that comprise a and Discrete Cosine Transform (DCT) to hide an image in an
header and a data block. The main data section in a particular audio file. The audio file format is not specified in the paper
frame contains two pieces: granule1 and granule2. If MP3 which is important for audio steganography because usually
file is of the mono-type, then each granule provides data for the compressed formats are preferred, and when compressed,
one channel, while if it is of the stereo-type, then each the secret information should not be lost. It was able to
granule provides data for two channels. Between the frames, achieve a hiding capacity of only 1.56% [6]. One recent
there may be space that does not contain any specific publication states about the insertion of secret bits in the sign
information, called ancillary data [2]. bits of the MP3 bit stream where the capacity is minute [7].
If the information is hidden in the main data segment, Since about the last two decades, steganography is being
there is no guaranty on its existence after the encoding combined with cryptography to enhance the security.
process. It is not possible to predict accurately, the loss However, using standard encryption techniques in addition to
during compression. It depends upon the encoder used and the steganographic algorithm increases the complexity. One
the compression algorithm involved. Therefore, data frames such proposal is found in [8], which uses AES encryption

978-1-7281-6324-6/20/$31.00 ©2020 IEEE 640

Authorized licensed use limited to: Carleton University. Downloaded on June 06,2021 at 05:58:01 UTC from IEEE Xplore. Restrictions apply.
technique and the MD5 hash function to improve the MP3 If the sample value is 0.87890619, its 32-bit floating-
steganography. Even in this paper, the capacity is very little. point representation is
If the audio file is a WAV file, then information 0 01111110 11000001111111111111111; bits in bold are
processing is not difficult. The authors in [9] proposed the the bits to modify.
audio steganography techniques in transform domain, which
improve the security and also enhance the capacity. But they Now suppose all the bits to be modified are changed to 1
have used WAV files in their implementation. If the secret (one extreme case),
information is an audio and the cover is any other file such as 0 01111110 11111111111111111111111; value increases
image [10], then the embedding is not complex. Even if the to 0.99993896.
secret audio is in MP3 format, it can be converted to WAV
format and then it can be hidden. But if this conversion is Then the 2nd bit is changed to 0 so that the value becomes
used for cover MP3 audio and then used in steganography 0.87108875
for embedding, the secret information cannot be retrieved if Similarly, if the sample value is 0.87500000, its 32-bit
the stego audio is converted back to MP3. Mohammed floating-point representation is
Hatem Ali Al-Hooti et al. [11] proposed a method to hide the
0 01111110 11000000000000000000000; bits in bold are
secret information in the WAV cover audio with logical
the bits to modify.
operators using random sample bits. The authors in [12]
generated a new audio channel to hide the secret data, which Now suppose all the bits to be modified are changed to 1
may increase the cover audio size. (the other extreme case),
A secure technique with less complexity and high 0 01111110 11111110000000000000000; value increases
capacity to hide a large amount of secret information is of to 0.99609375.
necessity, and one such method is proposed in this paper.
Then the 2nd bit is changed to 0 so that the value becomes
III. PROPOSED METHOD 0.87109375
To protect the secret information against MP3 Similarly, it can be shown that the change in the sample
compression and to achieve good capacity, the secret value is minimum by making the 2nd bit of the significand to
information is concealed in the main data, in duplicates. To 1 when all the bits to be modified are changed from 1 to 0.
overcome data loss, the audio samples with more than 50%
of the maximum sample value, are selected and the secret The secret information considered in this proposal is an
data is hidden in duplicates. Majority evaluation is done to audio file. The secret audio is transformed to the time-
approximate the hidden data while extracting. In MP3, frequency domain using Integer Wavelet Transform (IWT).
samples are stored in single precision (32-bit) floating-point This produces approximation and detailed coefficients of the
format. While reading, the samples are read in their native secret audio. IWT can be applied using lifting wavelets [13],
format, and the data is hidden in the significand field. The [14], [15]. Approximation coefficients have significant
sample values range between -1 and +1. features and using only these coefficients, entire secret
audio can be reconstructed. Thus, only approximation
A. Embedding Procedure coefficients are hidden in the cover. In the extraction
Up to 5 bits per sample can be hidden without any bit process, detailed coefficients are taken as zeros during
error. This improves the capacity. The 3rd, 4th, 5th, 6th and 7th inverse transformation. Hiding the frequency coefficients of
bit of the significand field from the most significant bit side the secret audio improves its security as it is not easily
is replaced by the secret bits. The positional weights of these detectable by eavesdroppers and the capacity is improved
bits (i.e., 1/8, 1/16, 1/32, 1/64, 1/128) add up to less than since only approximate coefficients are hidden. Capacity
0.25. Therefore, even on modifying all the bits to one, the and security do not go hand in hand. When the capacity is
change in the value will be less than 0.25. Since only the increased security goes down and vice-versa. The proposed
sample values with more than 50% of the maximum value method in this paper improves the capacity and the security
(i.e. more than 0.5) are considered for steganography, the 1st compared to the existing methods as discussed in the
bit of the significand is 1 in the normalized floating point ‘Results and discussion’ section. Built-in functions for
format except for the value 1. If the sample value is 1 then wavelet transformation and audio processing are available in
the stego value may increase beyond 1 and hence it is MATLAB and those function are used in this proposal for
excluded from embedding. Here the normalization is done experimentation.
such that the bit to the right of the radix point is one in the
binary representation of the sample value. This is done by This method is referred to as MP3AS. The block
moving the radix point and adjusting the exponent. Dynamic diagram of this proposed method is shown in Fig. 1. The
adjustment is made such that the sample value changes by embedding and extracting algorithms are shown in Fig. 2
only the minimum value. To achieve this, if all the bits to be and Fig. 3 respectively.
modified are changed from 0 to 1, then the 2nd bit of the B. Extracting Procedure
significand is made 0. If all the bits to be modified are The hidden secret audio data can be extracted by
changed from 1 to 0, then the 2nd bit of the significand is performing the operations of embedding in the reverse order.
made 1. In other cases, no adjustment is done. Two extreme The algorithm is shown in Fig. 3
cases are illustrated in the example below.

641

Authorized licensed use limited to: Carleton University. Downloaded on June 06,2021 at 05:58:01 UTC from IEEE Xplore. Restrictions apply.
 IV. RESULTS AND DISCUSSION TABLE I. SNR AND SPCC VALUES FOR DIFFERENT SECRET SAMPLE SIZES
WITHOUT DYNAMIC ADJUSTMENT.
To write MP3 files LAME 3.99.5 encoder tool is used. No. of SNR in dB SPCC
Results show that high capacity with reasonably good Cover
secret
performance metrics are obtained compared to the existing audio Extracted Extracted
bits per Stego Stego
MP3 steganography techniques. If the header information is samples Secret Secret
sample
used for hiding the information, only 5 to 6 bits per frame
can be hidden. Compared to that, the proposed technique 2553062 3(29905
40 32 0.9025 0.8041
3 out of samples)
gives very high capacity with good security. The proposed which
method is evaluated using the metrics SNR and SPCC. 159496 4(39874
32 31 0.8015 0.795
with samples)
SNR indicates the deviation between the original and the
value
modified signals. It is given by (1) [16]. more 5(49842
29 30 0.7753 0.7823
than 0.5 samples)
SNR = 10 (1)
TABLE II. SNR AND SPCC VALUES FOR DIFFERENT SECRET SAMPLE SIZES WITH
where DYNAMIC ADJUSTMENT.
No. of SNR in dB SPCC
x and y are the original and the modified signals Cover
secret
respectively. audio Extracted Extracted
bits per Stego Stego
samples Secret Secret
sample
N is the number of signal samples
2553062 3(29905
xi is the value of the ith sample in x 70 32 0.9725 0.8041
3 out of samples)
which
yi is the value of the ith sample in y. 4(39874
159496 67 31 0.9225 0.795
with samples)
A larger SNR value indicates a better quality. The
recommended SNR for audio signal is above 30dB. value
more 5(49842
62 30 0.9000 0.7823
SPCC is a metric used for measuring the quality of an than 0.5 samples)
audio signal which depends on the correlation of samples.
The high value of SPCC indicates good quality. Its range is
between 0 and 1 and it is given by (2) [16]. V. PERFORMANCE AGAINST ATTACKS
The common attacks on stego audio signals are the
(2) compression, filtering, resampling, normalization etc. [17].
The cover audio samples with the secret audio bits are
where written as MP3 audio file and it is possible to retrieve the
secret audio signal. Therefore, this technique withstands
x and y are the input and the output signals respectively. MP3 compression. Since only the audio samples with more
and are the average of the input and the output than 50% of the maximum sample value, are selected for
signals respectively. embedding, the possibility of losing these samples during
compression is minimum. If all the samples are used for
The SNR and SPCC are computed for constant cover embedding, then the extracted secret data is not audible
sample and different secret bits per sample. Table 1 and properly as some of the samples might have been lost when
Table 2 show the values for these metrics without and with the stego audio file is compressed. It is illustrated in the
the use of dynamic adjustment respectively. In these tables, results shown in Table 3. The SNR and SPCC values for the
the number in the brackets in the second column indicates extracted secret audio are very less, which indicate that the
the number of secret audio samples. Comparing Table 1 and audio cannot be audible.
Table 2, with the use of dynamic adjustment, SNR and
SPCC values are well improved. This proves that the hidden Table 4 shows the performance against other common
information is imperceptible. The plots of the cover and the attacks. Bit Error Rate (BER) is the percentage of the bit
secret audio are shown in Fig. 4 and Fig. 5. Fig. 6 and Fig. 7 error. It is found that the BER gives the appropriate
display the plots of the stego and the extracted secret audio measurement when the technique is tested for robustness. If
for which the metrics are shown in Table 2. the stego audio file is cropped at the beginning and ending,
it will not affect the hidden data because MP3 files have
The values in Table 2 indicate that with 5 secret bits per some amount of silence period at both the beginning and the
cover sample, the metrics of the stego and the extracted end. It is cropped 10% in time (at the beginning and ending)
secret audio are well within the range. The SNR is above 30 and by 1dB in amplitude. Then the bit error rate of the
dB and the SPCC is close to one for both stego and extracted secret audio is 12%. If the cropping is increased
extracted secret audio. When the number of secret bits per further, the secret audio cannot be retrieved. The secret
cover sample is increased further, the secret audio cannot be audio can be extracted without high bit-error when filtering
extracted properly. Good quality stego object is an (both low pass and high pass) with 1 kHz cut off frequency.
indication of secure steganography system. The robustness Resampling with a difference in the sampling rate of 0.5
of the proposed method is explained in the following section. kHz and normalisation up to plus or minus 10% gain will
not affect the hidden data. Outside this range, the extracted
audio will be distorted

642

Authorized licensed use limited to: Carleton University. Downloaded on June 06,2021 at 05:58:01 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1. Block diagram – MP3AS (Embedding)

Algorithm: Embed-(MP3AS) embeds the secret audio in the cover audio.

Input: Cover audio C.mp3, Secret audio S.mp3
Output: Stego audio, G.mp3.
Method:
1. Read cover audio C and secret audio S and convert the secret audio to .wav format
2. LS Å liftwave ( ‘haar’, 'Int2Int' )
3. [CAs, CDs] Å lwt(double(S),LS) // Obtain IWT of the secret audio to get approximation (CAs) and detail
(CDs) coefficients
4. CAs_bin Å dectobin(CAs)
5. N Å No. of bits in CAs_bin
6. Nbits Å dectobin(N) // Convert approximate coefficients to binary and obtain the no. of secret bits N.
Convert N to binary
7. Copy Nbits to an array D. Duplicate CAs_bin four times and append to D.
8. for i= 1 to length(C) do
while(N>0) do
if (C(i) >0.5 ) then
B Å floattobin(C(i))
B(3:7) Å secretbits from D
C(i) Å bintofloat(B)
end if
N Å N-5
end while
end for // Hides the secret bit array D in the 3rd, 4th, 5th, 6th, and 7th bits of the significand of the cover samples
whose value is greater than 0.5.
9. audiowrite(C, ‘G.mp3’) // Write stego audio file
10. return stego audio G.
Fig. 2. MP3AS embedding algorithm

643

Authorized licensed use limited to: Carleton University. Downloaded on June 06,2021 at 05:58:01 UTC from IEEE Xplore. Restrictions apply.
Algorithm: Extract-(MP3AS). Extracts the secret audio from the stego audio.
Input: Stego audio G.mp3
Output: Secret audio S_extracted.mp3.
Method:
1. Read stego audio G
2. Convert elements of G with a value greater than 0.5, to 32-bit floating point binary. Extract 3rd, 4th, 5th, 6th
and 7th bits of significand. The first N bits will give the no. of secret bits of approximate coefficients.
3. CAs_bits Å bintodec(N)
4. Extract four times CAs_bits from the elements of G.
5. Perform the majority evaluation in the four groups of the CAs_bits and obtain the approximate coefficients,
CAs by converting to decimal.
6. S_extracted Å ilwt (CAs, 0, LS) // Obtain the secret audio by taking the inverse transform considering
detailed coefficients as zero.
7. return secret audio S_extracted.mp3. // Write to a file and return
Fig. 3. MP3AS extracting algorithm

Fig. 4. Cover audio Fig. 5. Secret audio

Fig. 6. Stego audio Fig. 7. Extracted secret audio

644

Authorized licensed use limited to: Carleton University. Downloaded on June 06,2021 at 05:58:01 UTC from IEEE Xplore. Restrictions apply.
 TABLE III. SNR AND SPCC VALUES FOR DIFFERENT SECRET SAMPLE SIZES
WHEN ALL THE COVER SAMPLES ARE USED FOR EMBEDDING.
TABLE IV. PERFORMANCE AGAINST COMMON ATTACKS.

No. of SNR in dB SPCC Extracted audio performance

Cover
audio
secret Attack Type metrics
bits per Extracted Extracted
samples Stego
Secret
Stego
Secret
SNR in dB SPCC BER (%)
sample
Cropping 25 0.7581 12
3(29905
52 23 0.9125 0.5045 Low pass filtering 29 0.8030 8.5
samples)
25530623 4(39874 High pass filtering 28 0.8015 9
43 18 0.9015 0.3215
samples samples) Resampling 30 0.8875 6
5(49842 Normalization 32 0.8876 5
35 10 0.8953 0.156
samples)

and Communication Networks ,2018, pp. 1-21.

VI. CONCLUSION [7] M. Bazyar and R. Sudirman: 'A New Data Embedding Method for
Audio signals can be efficiently used for steganography MPEG Layer III Audio Steganography', Jurnal Teknologi, 2016, 78
(7-5), pp. 67-73.
because of its ease in storage and transmission. This paper
[8] R. Indrayani, H. A. Nugroho, R. Hidayat and I. Pratama: 'Increasing
proposes an MP3 audio steganography technique which is
the Security of MP3 Steganography Using AES Encryption and MD5
challenging since the MP3 format is complex. The results of Hash Function'. Proc. 2nd International Conference on Science and
this proposal outperform, with respect to capacity and Technology-Computer. Yogyakarta, Indonesia, October 2016.
security. It withstands the common attacks such as [9] M. Z. Muzaffar, 'Compressed Sensing for Security and Payload
compression, cropping, filtering, resampling and Enhancement in Digital Audio Steganography', Journal of
normalization demonstrating its robustness. MATLAB 8.2 is Information Hiding and Multimedia Signal Processing, 2018, 9 (6),
pp. 1506-1518.
used for implementation.
[10] S. W. Naing and T. M. Naing: 'Secret Audio Messages Hiding in
REFERENCES Images'. Proc. International Conference on Big Data Analysis and
Deep Learning Applications. Miyazaki, Japan, 2018, pp 215-223.
[1] A. Mohammed Salem, I. Subariah, S. Ghazali, Z. Akram and A. [11] M. H. A. Al-Hooti, T. Ahmad and S. Djanali: 'Developing audio data
Adamu: 'Exploring the Challenges of MP3 Audio Steganography'. hiding scheme using random sample bits with logical operators',
Proc. International Conference on Advanced Computer Science Indonesian Journal of Electrical Engineering and Computer Science',
Applications and Technologies, Kuching, December 2013, pp 156- 2019, 13 (1), pp. 147-154.
161
[12] S. Z. Khan and Y. Ullah: 'Audio Steganography by Additional
[2] Z. Mikhail, L. William and C. Robert: 'MP3 Files as a Steganography Channel'. Proc: ICCD - Recent Developments in Intelligent
Medium'. Proc. Second annual conference on Research in Computing, Communication and Devices, 2017.
information technology, Orlando, Florida, USA, October 2013, pp
[13] L. Baiying, S. Ing Yann, Z. Feng, L. Zhen and L. Haijun: 'A Robust
23-28.
Audio Watermarking Scheme based on Lifting Wavelet Transform
[3] Y. Diqun, W. Rangding, Y. Xianmin and Z. Jie: 'Steganography for and Singular Value Decomposition', Signal Processing,, 2012, 92, pp.
MP3 Audio by Exploiting the Rule of Window Switching', 1985-2001.
Computers and Security, 2012, 31, pp. 704-716.
[14] D. Fatiha, A. Beghdad, M. Karim Abed and H. Hamam:
[4] C. Bingwei and Z. Jiying: 'An Adaptive and Secure Audio 'Comparative Study of Digital Audio Steganography Techniques',
Watermarking Algorithm Robust to MP3 Compression', Proc. EURASIP Journal on Audio, Speech, and Music Processing, 2012.
International Instrumentation and Measurement Technology
[15] S. Hemalatha, A. U. Dinesh and A. Renuka: 'Audio data hiding
Conference. Singapore, May 2009.
technique using integer wavelet transform', International Journal of
[5] Y. Bailong, W. Penghui, J. Yaque and M. Jing: 'Lossless and Secure Electronic Security and Digital Forensics, 2016, 8 (2), pp. 131-146.
Watermarking Scheme in MP3 Audio by Modifying Redundant Bit in
[16] B. L. Dora M. and M. A. Juan M: 'Highly Transparent Steganography
the Frames'. Proc. International Conference on Information
Model of Speech Signals using Efficient Wavelet Masking', Expert
Management, Innovation Management and Industrial Engineering.
Systems with Applications, 2012, 39 (10), pp. 9141-9149.
Xi'an, China, November 2013.
[17] M. Nutzinger: 'Real-time Attacks on Audio Steganography', Journal
[6] B. G. Banik and S. K. Bandyopadhyay: 'Blind Key Based Attack
of Information Hiding and Multimedia Signal Processing, 2012, 3
Resistant Audio Steganography Using Cocktail Party Effect', Security
(1), pp. 47-65.

645

Authorized licensed use limited to: Carleton University. Downloaded on June 06,2021 at 05:58:01 UTC from IEEE Xplore. Restrictions apply.