There is an apparently a provably secure cipher that was proposed by Diffie, but enhanced by another cryptographer. The scheme, which was mentioned in Applied Cryptography, works like this:

1. Measure the length of the plain-text, $n$.
2. Multiply it by $128$.
3. Generate this much ($128·n$ bytes of) real random data and split it out into 128 byte-arrays of length each equal to the plain-text. This can be thought as a two dimensional array:
   1. One of the indices gives the sequence number ($0\dots 127$).
   2. One of the indices gives the position in the sequence, $0 \dots n-1$.
4. Use a 128-bit key to choose which of these streams to XOR together. Each bit of the key corresponds to "yes/no" on whether to use particular sequence. All the selected sequences are XORed together to make a single keystream, $K$.
5. Compute $P \oplus K$ to give the cipher-text $C$.
6. Serialize the two dimensional array and append it to the cipher text.
7. Send the whole package to Bob, who can then decrypt by de-serializing the matrix and selecting the same rows.

Apparently, this scheme is completely secure. The attacker has to examine every possible combination of sequences ($2^{127}$ on average) in order to break the encryption scheme.

What is the proof of this? I can't find the paper that discusses this anywhere.