Perl is notorious for being able to parse stuff which looks like garbage, there's a whole category for Obfuscated code on PerlMonks. So let's hope your Perl programmers do this in less than 20% of their files ;)

For the general task of classifying data, there's AI::NaiveBayes and AI::Categorizer. They both need some adaption to parse text into the categories "Perl source code" and "garbage". I would guess that you get 80% accuracy with a filter based on the regular expressions presented by other monks, so only if this fails, training a Bayesian might be an alternative.