http://www.cs.unc.edu/~fabian/papers/foniks-oak11.pdf
edit: this paper is referenced in fact by the paper submitted here.
From a bandwidth perspective, such CBR encodings are either wasteful or low quality for high motion scenes—or both. So it makes sense that Netflix has chosen a VBR system, but does have this privacy caveat.
To avoid that issue you just artificially make multiple movies have the same length, meaning do some padding to round up to the next e.g. 10 minutes (so if the movie is 1:48:23 it becomes 1:50:00), to do so the movie keeps buffering in the background (some random audiovisual noise).
Maybe netflix should have a "Fully safe" mode where it uses CBR instead of VBR so the user knows the trade off (slower/heavier buffering)
Different bit rates on the random audiovisual noise (white noise would be uncompress-able, thus weight more, pure black would be too compress-able, etc).
Some possible counter-measures are discussed in the conclusion:
"To that end, we believe that Netflix could defend against passive traffic analysis by ensuring that the byte-range portion of the HTTP GETs sent by the browser do not perfectly align with individual video segment boundaries. For instance, the browser could average the size of several consecutive segments and send HTTP GETs for this average size. As an alternative approach, the browser could randomly combine consecutive segments and send HTTP GETs for the combined video data. Designing obfuscation techniques for VBR DASH streams that do not degrade video quality remains a potential area for future research."
edit: clarity
Stepping back a bit, although this paper is definitely valuable, it isn't that startling, because we already know that encrypted communications are vulnerable to passive attacks when the contents are predictable. It's a good reminder that "vanilla" encryption isn't necessarily the best way to protect privacy when the attacker can simply guess what we're transmitting because the search space is so small; in this case, it's easy to compare the length of what is being transmitted against a corpus -- and bam. There's only ~42k entries...
Entropy entropy entropy. It is your friend. Just so happens that VBR and DASH weren't designed to increase entropy when transmitting segments.