I have seen lot's of platforms these days implement stringent policies as far as music copyright detection. Fortunately, there are still ways to get around these systems.
Avoid Audio Copyright
How do many online platforms "detect" songs?
All of the ingredients of a sound are determined by the sum of all of the different-frequency sound waves in the area. To "fingerprint" sound, you can run a little math tool called an "FFT" which convert sound information into the frequency domain. Think of this as a bar-chart expressing how much of each frequency there is. Each sound has a unique bar chart, and as a song plays, there is a pattern to it's time-dependent bar chart.
How do we get around this?
If we are able to distort this bar chart enough, then the computer systems cannot recognize our music. We have to be careful though, because if we distort it in the wrong ways, we too will identify the music!
Okay, how do we distort the music but not?
One way is we can follow along our music, adding noise of varying frequencies as we go, throwing off the sound of the song. This will definitely help throw the copyright systems off our trail, but it will sound pretty jank. Another way is we can run an adversarial AI to identify what portions of songs contribute to the most *recognizable* parts of each portion of music, and we can then just perturb the remaining parts of music.
This process is actually able to produce quality reproductions of songs, which don't sound audibly different to the untrained ear, yet bypass music copyright systems.
Full Implementation Of This
This was studied in-depth in a study: https://arxiv.org/pdf/1906.07153.pdf
There are a variety of implementations online, but hopefully this inspires some people on here to fork some repos and use this for some clever blackhat monetization.
Avoid Audio Copyright
How do many online platforms "detect" songs?
All of the ingredients of a sound are determined by the sum of all of the different-frequency sound waves in the area. To "fingerprint" sound, you can run a little math tool called an "FFT" which convert sound information into the frequency domain. Think of this as a bar-chart expressing how much of each frequency there is. Each sound has a unique bar chart, and as a song plays, there is a pattern to it's time-dependent bar chart.
How do we get around this?
If we are able to distort this bar chart enough, then the computer systems cannot recognize our music. We have to be careful though, because if we distort it in the wrong ways, we too will identify the music!
Okay, how do we distort the music but not?
One way is we can follow along our music, adding noise of varying frequencies as we go, throwing off the sound of the song. This will definitely help throw the copyright systems off our trail, but it will sound pretty jank. Another way is we can run an adversarial AI to identify what portions of songs contribute to the most *recognizable* parts of each portion of music, and we can then just perturb the remaining parts of music.
This process is actually able to produce quality reproductions of songs, which don't sound audibly different to the untrained ear, yet bypass music copyright systems.
Full Implementation Of This
This was studied in-depth in a study: https://arxiv.org/pdf/1906.07153.pdf
There are a variety of implementations online, but hopefully this inspires some people on here to fork some repos and use this for some clever blackhat monetization.