Made me remember "Little Brother" by Cory Doctorow. The book described Paranoid Linux in this way: “Paranoid Linux is an operating system that assumes that its operator is under assault from the government (it was intended for use by Chinese and Syrian dissidents), and it does everything it can to keep your communications and documents a secret. It even throws up a bunch of “chaff” communications that are supposed to disguise the fact that you’re doing anything covert. So while you’re receiving a political message one character at a time, Paranoid Linux is pretending to surf the Web and fill in questionnaires and flirt in chat-rooms. Meanwhile, one in every five hundred characters you receive is your real message, a needle buried in a huge haystack."
That approach may be easier to defeat nowadays with DNNs - they would be extremely good at getting the noise out of the signal, so I am not sure it's enough to counter any nation state capabilities who really want to know what you are doing.
Why do you think so? If the patterns are relatively predictable and constant (which would be the case if they are programmed to issue some specific noise on a regular basis), Machine Learning will have very good accuracy and distinguishing what is signal vs what is noise. Much better and faster than humans, that's for sure.
