roneill / cs4500 Goto Github PK

Team Steganosaurus
==================

Team Members
------------
* Bob O'Neill ([email protected])
* Matt Garnes ([email protected])
* George Proctor ([email protected])
* Daniel Bostwick ([email protected])

Usage
-----
To run our software on the CCIS Linux machines, simply cd
into the /CS4500/Stegan/bin/ directory and run the stegan executable with
the correct arguments. Following the specifications given for the
semester project software, the available commands are:

Encode a payload into a container audio file and write the result to
trojan > stegan --encode container payload trojan

Decode the data hidden in trojan into payload
>  stegan --decode trojan payload

In any place that you specify the trojan or the container audio file
path, including the output path for encode, you can use an mp3 or wave
file. Our software will make the necessary conversions using the Lame
executable.

Controlling Steganography
-------------------------

There is a value in fft_encode.py called base_power_value that
controls the steganography of the payload. This value is currently set
to 1200000. This value was choosen because it offers a good mix of
steganographc quality and reliablity of payload retrieval. A value of
300000 or so will greatly increase steganographic quality at the
expense of payload retrieval reliability.

Algorithm Explanation 
---------------------
First, we break the container audio file into chunks, each chunk being
one second of audio on one channel (44100 samples). We calculate the
"density" of the encoding, defined as the number of bytes that will be
encoded into each chunk of audio. Then, for each chunk, we take an fft
of the audio to find the power of each frequency. For every payload
byte that will be encoded into the chunk, we look at the power of the
frequencies in a certain range (for example 15000 Hz to 15255 Hz) and
increase the power of the frequency that will represent the byte to be
encoded. The power of the frequency is determined by starting from a
base value (determined through experimentation), increasing the power
until the value can be reliably decoded. Each frequency represents a
byte value from 1 - 256. To encode further bytes in the same chunk,
this process is repeated using more ranges of frequencies lower than
before. Using this strategy, we also encode the length of the payload
at the beginning of the file to aid in decoding.

To decode, we chunk the input audio file in the same way as before. We
first decode the payload size and with that compute the density of the
data. For each chunk we take an fft and find the power of each
frequency. Using the computed density value, we know exactly which
frequency ranges we expect data to be encoded, so we find the
strongest frequency in each range and write the byte that this
frequency represents. In this way we can reconstruct the data from the
original payload.

Limitations/Known Issues When the Output is Converted to MP3
------------------------------------------------------------
1. We have found that our software performs
best with quieter audio. The ideal candidate is acoustic music, such
as the example wav file that you gave us, Hewlett.wav. When we have
tested our software with loud input or input with a large variety of
frequencies present (such as electronic music), the software will
attempt to compensate for the loud frequencies already present by
increasing the volume. The result is a very un-steganographic
container file filled with loud beeps and tones.

2. Our software is technically able to operate on payloads with a size
up to 0.1% of the carrier audio file size. However, since our
algorithm broadens the range of frequencies that will be used to
encode the data the larger the payload is, this will result in less
steganographic results. In addition, since this broader range of
frequencies starts to move into ranges that are more likely to have
existing music in them, the reliability of decoding suffers. Some of
the data is not able to be recovered. So, in short, the smaller the
payload, the more accurate and steganographic the output will be. A
user may be interested in compression payloads before encoding them
using our software. This will allow the user to encode a much larger
amount of data, although they must be careful that nothing is lost,
making decompression of the result impossible.

3. The bit rate of the MP3 conversion has a large impact on the
accuracy of decoded payload data. If the bit rate is 320 kbps, the
accuracy is much higher than a 128kbps or 192kbps MP3 file. We suggest
that 320 kbps conversion is used for the best results.

Third Party Software Used
-------------------------
1. The Lame executable
2. Python 2.6 and several of its built-in modules (numpy, scipy, wave)