Implement Adaptive Padding or some variant and measure overhead vs accuracy
As a defense against Website Traffic Fingerprinting, we should implement a tunable cover traffic defense that we could set from the consensus with a value dependent upon available Guard bandwidth relative to Exit capacity.
My favorite from the research literature is http://freehaven.net/anonbib/cache/ShWa-Timing06.pdf, because it appears to be tunable in this fashion.
The "BUFLO" variant proposed by this paper is better specified, but it's not clear it actually performs better for a given overhead quantity: http://www.cs.sunysb.edu/~xcai/fp.pdf
This is likely a research task. People who attempt it should also read http://www.raid-symposium.org/raid99/PAPERS/Axelsson.pdf (Slides: http://www.cse.psu.edu/~tjaeger/cse543-f06/presents/Kiran_baserate.pdf)