Compressed archives are probably the very worst format for trying to insert a subchannel. The problem here is that almost every bit change in an archive has an effect on many bits in the unpacked contents, and in a way that depends on the whole archive contents. Changing a bit or two at random is extremely likely to produce unpacked files that have invalid file formats (or just corrupt archives). This is easy for an attacker to notice. About the only place a few bits of subtext might be located is by taking advantage of the error-correcting codes (ECC) some archive formats use. You could introduce an occasional "error" in archives of the type the ECCs would correct upon unpacking. One trick would be to make sure that archives with subtexts do not have many more errors than archives without subtexts (which means introducing random "errors" to all transmitted archives that an attacker might intercept).

Natural language text. Natural language is extremely free-form, and thus is an excellent format in which to embed a subchannel. Normal texts contain all sorts of spacing variations, word-choices, types, and other "random" features. But then, a too-obvious subchannel encoding strategy is easy to detect. Sure, people make typos, but not in uniformity in every third word. Too much pattern in the "random" variations is easy for a machine scan, or a human reader, to identify as a probable subtext.