Check this out:

http://sourceforge.net/projects/stlth/

It's like Truecrypt but based on dm-crypt, GPL and supports unlimited numbers of hidden volumes.
That's real plausable deniability, unlike Truecrypt.

My Idea is to make this system completely secure, no matter what resources are spent cracking it. At least in theory. ^^

Are you familiar with the cryptsetup program? When used (with or without LUKS) it doesn't copy anything and definitely doesn't write unencrypted data to the harddrive. Instead it creates a mapping, a new device in /dev/mapper/, which can be accessed to read and write to the file. This means that leaving traces of unencrypted data is not a problem.

Acually I did some more research. Check this out (read the comments):
http://it.slashdot.org/story/09/04/30/201222/Forensics-Tool-Finds-Headerless-Encrypted-Files
It's about the same thing I'm working on, and it seems that my tool will work. ^^

Let me give an example:
I have one (1) 100M file that looks random (could just as well be a Truecrypt volume). When I open it with my program I say that my data starts 20M into the file, and enter my password. Then `dumpe2fs` says that the filesystem size is 30M. This means that my data is written between Mbytes 20-50. I can make the filesystem larger and add files if I wish to, but everytime I unmount the volume I use resize2fs to shrink the fs as much as possible.
I also have a "secret" volume between Mbyte 60-80. But can I claim that this is just random data, that can be used to expand my first volume?