sunzhen1994 / deanonymizing_nucleus_sampling Goto Github PK

This supplementary material contains the follow files:

NSS_reddit_asoiaf.pkl: Nucleus size series(NSS) of 500 2700-word sequences from the reddit-asoiaf dataset. The data contains 500 rows; each row is a NSS with length 2700.

NSS_reddit_electronic_cigarette.pkl: NSS of 500 2700-word sequences from the reddit-electronic-cigarette dataset. Same format as NSS_reddit_asoiaf.pkl.

NSS_reddit_india.pkl: NSS of 500 2700-word sequences from the reddit-india dataset. Same format as NSS_reddit_asoiaf.pkl.

NSS_reddit_OkCupid.pkl: NSS of 500 2700-word sequences from the reddit-OkCupid dataset. Same format as NSS_reddit_asoiaf.pkl.

NSS_reddit_Random_Acts_Of_Amazon.pkl: NSS of 500 2700-word sequences from the reddit-Random-Acts-Of-Amazon dataset. Same format as NSS_reddit_asoiaf.pkl.

NSS_reddit_sports.pkl: NSS of 4000 2700-word sequences from the reddit-sports dataset. The data contains 4000 rows; each row is a NSS with length 2700.

NSS_vs_trace.pkl: 1473 (NSS, trace) paris from the reddit-sports dataset. Traces are processed to remove noise and outliers. The data contains 1473 rows; each row is a (NSS, trace) pair.

silkroad_NSS_vs_trace.pkl: (user, NSS, trace) tuples of 41 silkroad users. The data contains 41 rows; each row is a (user, NSS, trace) tuple.

All data files are in '.pkl' format. Please use the Python pickle module to load the data.

plot_silkroad.py: A Python script shows how to match NSSes against traces using the above silkroad data. Results are plotted in a confusion matrix of NSS vs. trace. To run the script: "python3 plot_silkroad.py".



mitigation.diff: contains our migitigation code, given as a ``git diff'' from Hugging Face Transformers repository code, at commit a615499076a. To apply this patch, perform:

====
git clone https://github.com/huggingface/transformers.git
cd transformers
git checkout a615499076a67dceb8907ecdf8eadaff04bb8d6a
git apply <path-to-mitigations.diff>
====