But the metadata is a form of structured data and as you state is semi known plaintext; i.e. the structure is known, and to a lesser extent also the data (some

But the metadata is a form of structured data and as you state is semi known plaintext; i.e. the structure is known, and to a lesser extent also the data (some fields have a known or limited range of possibilities). Aside from the headers the compressed data it self often has structure; take the Lempel Ziv class of dictionary encoders rely on repeating data to compress. It is just this fact that it is repeating means that you can guess with much higher probability than normal what certain bytes will _not_ be (because longest words that match the dictionary are chosen to tokenised to maximise compression); i.e. bytes that _don't_ match a word suffix in the dictionary will restart the search for a new matching word / token pair. Having said that the plaintext itself is almost never random; but the key thing does the attacker have any crib that might be used to have a good first guess that can reduce the amount of work required. So which is more guessable; the semi-structure, at the byte level, of compressed data, or the possible semi-structure of the original plain text? If you are dealing with a protocol that specifies compression (and in particular which compression method) you may have given away part of the game. One way is to "bump up" the entropy and add "chaff" to the compressed stream before encryption; i.e. add some entropy but less than what would be less than the amount saved through compression. My gut feel is that the efficacy of this would vary depending on plaintext, compression method, and encryption method. You also run the risk of side channel analysis via CPU, RAM, power usage etc.