It's out of similar considerations that some time ago I wrote dupseek. Even for multiple files, just using parts of file contents instead of digests works quite well (and rules out even the very small risk of collisions).