This is a a golang implementation of the kNN website fingerprinting attack and feature extractor by Wang et al..
There are three versions:
knn.origwhich produces identical output as Wang's implementation in the attacks collection.knn.fixedthat fixes some bugs in feature extraction, removes unnecessary features for Tor, and fixes a bug in the weight learning algorithm. The feature extraction is (to the best of my knowledge) identical to Table 3.1 in Wang's PhD thesis.go-knnthat provides an optmized go-knn version with multi-core and sub-folder support for work.
Initial tests show no meaningful differences between the three versions beyond speed due to removing unnecessary features for Tor from the feature extraction. The port and bugfixes were a way for me to better understand the attack.
Download Wang et al.'s cell traces
for their USENIX 2014 paper with 100x90 + 900 traces. Extract the traces, creating the batch folder.
Extract the original features:
$ go run cmd/feat.orig/fextractor.go -sites 100 -instances 90 -open 9000
2016/04/12 11:31:15 starting parsing...
2016/04/12 11:32:27 done parsing (100 sites, 90 instances, 9000 open world, folder "batch/", suffix "f")
Extract the fixed features:
$ go runcmd/feat.fixed/fextractor.go -sites 100 -instances 90 -open 9000
2016/04/12 11:30:29 starting parsing...
2016/04/12 11:30:46 done parsing (100 sites, 90 instances, 9000 open world, folder "batch/", suffix "s")
Run the original attack:
$ go run cmd/knn.orig/knn.orig.go
2016/04/12 11:32:49 loaded instances: main
2016/04/12 11:32:50 loaded instances: training
2016/04/12 11:32:52 loaded instances: testing
2016/04/12 11:32:55 loaded instances: open
2016/04/12 11:32:55 starting to learn distance...
distance... 8999 (0-9000)
2016/04/12 11:47:18 finished
2016/04/12 11:47:18 started computing accuracy...
accuracy... 17999 (0-18000)
2016/04/12 12:43:11 finished
2016/04/12 12:43:11 Accuracy: 0.870667 0.989889
Run the fixed attack:
$ go run cmd/knn.fixed/knn.fixed.go
2016/04/12 11:32:45 loaded instances: main
2016/04/12 11:32:46 loaded instances: training
2016/04/12 11:32:46 loaded instances: testing
2016/04/12 11:32:46 loaded instances: open
2016/04/12 11:32:46 starting to learn distance...
distance... 8999 (0-9000)
2016/04/12 11:36:50 finished
2016/04/12 11:36:50 started computing accuracy...
accuracy... 17999 (0-18000)
2016/04/12 11:54:32 finished
2016/04/12 11:54:32 Accuracy: 0.872556 0.989889
As far as a straight port from source code can be licensed by me (probably not at all), the code is licensed under GPLv3.
This work is part of the HOT research project, funded by the Swedish Internet Fund.