Fingerprint (computing)

To serve its intended purposes, a fingerprinting algorithm must be able to capture the identity of a file with virtual certainty. In other words, the probability of a collision — two files yielding the same fingerprint — must be negligible, compared to the probability of other unavoidable causes of fatal errors (such as the system being destroyed by war or by a meteorite): say, 10⁻²⁰ or less.

This requirement is somewhat similar to that of a checksum function, but is much more stringent. To detect accidental data corruption or transmission errors, it is sufficient that the checksums of the original file and any corrupted version will differ with near certainty, given some statistical model for the errors. In typical situations, this goal is easily achieved with 16- or 32-bit checksums. In contrast, file fingerprints need to be at least 64-bit long to guarantee virtual uniqueness in large file systems (see birthday attack).

When proving the above requirement, one must take into account that files are generated by highly non-random processes that create complicated dependencies among files. For instance, in a typical business network, one usually finds many pairs or clusters of documents that differ only by minor edits or other slight modifications. A good fingerprinting algorithm must ensure that such "natural" processes generate distinct fingerprints, with the desired level of certainty.

Computer files are often combined in various ways, such as concatenation (as in archive files) or symbolic inclusion (as with the C preprocessor's #include directive). Some fingerprinting algorithms allow the fingerprint of a composite file to be computed from the fingerprints of its constituent parts. This "compounding" property may be useful in some applications, such as detecting when a program needs to be recompiled.

Rabin's fingerprinting algorithm[7] is the prototype of the class. It is fast and easy to implement, allows compounding, and comes with a mathematically precise analysis of the probability of collision. Namely, the probability of two strings r and s yielding the same w-bit fingerprint does not exceed max(|r|,|s|)/2^w-1, where |r| denotes the length of r in bits. The algorithm requires the previous choice of a w-bit internal "key", and this guarantee holds as long as the strings r and s are chosen without knowledge of the key.

Rabin's method is not secure against malicious attacks. An adversarial agent can easily discover the key and use it to modify files without changing their fingerprint.

Mainstream cryptographic grade hash functions generally can serve as high-quality fingerprint functions, are subject to intense scrutiny from cryptanalysts, and have the advantage that they are believed to be safe against malicious attacks.

A drawback of cryptographic hash algorithms such as MD5 and SHA is that they take considerably longer to execute than Rabin's fingerprint algorithm. They also lack proven guarantees on the collision probability. Some of these algorithms, notably MD5, are no longer recommended for secure fingerprinting. They are still useful for error checking, where purposeful data tampering isn't a primary concern.