Two videos, X and Y, are uploaded to youtube. X is a video of a completely featureless white screen. Y is a video of a jungle canopy in the midst of a storm.

Two videos, X and Y, are uploaded to youtube. X is a video of a completely featureless white screen. Y is a video of a jungle canopy in the midst of a storm. At some point in both videos, one pixel changes color slightly. Which video do you think it will be easier to spot the change in? Of course, the amount of information that can be transmitted in the color change of one pixel is ridiculously small, so in a real life example more pixels (or perhaps some other data in the video) would need to be used to embed the message, but the video size can grow along with the size of the hidden message. As for "anomaly detection", the thresholds at which such detectors function have to be tuned in such a way that they don't give too many false positives to make them useless. And they're not magic. They can only detect certain types of anomalies, not any an all past or future steganographic techniques that could conceivably be used to hide the message. Steganographic techniques can and have been designed to mimic expected statistical profiles. Take a look, for instance, at Peter Wayner's work on Mimic Functions: http://en.wikipedia.org/wiki/Mimic_function http://en.wikipedia.org/wiki/Mimic_function The other thing I should note is that even if it is (theoretically) possible to detect a message hidden via steganography, the cost of doing so goes up as the amount of data the message is hidden in increases. Even detecting a message hidden with even the simplest steganographic technique will be much higher than detecting the use of bare encryption, which is already out in the open. So widespread use of steganography in large datastreams like youtube videos and Skype will create a huge computational burden on the snoopers attempting steganalysis.