HE Spamfilter

SNUCSE 2021 "Intelligent Computing System Design Project"

• Hyesun Kwak
• Myeonghwan Ahn
• Dongwon Lee

abstract

• Naïve Bayesian spam filtering
• with Homomorphic Encryption (using Lattigo implementation of CKKS)
• with BERT embedding (using huggingface/tokenizers)
• on enron dataset from MWiechmann/enron_spam_data
• using gin web framework

tested on

• client (both sender & receiver)
  • MacBook Air (M1, 2020)
  • macOS 12.0.1 (21A559)
  • Docker Desktop 4.3.0 (71786)
    • go version go1.13.8 linux/arm64
• server
  • intel Xeon Platinum 8268
  • ~200 GiB memory
  • Ubuntu 20.04.3 LTS
    • go version go1.13.8 linux/amd64
• common
  • Lattigo v2.3.0
  • huggingface/tokenizers v0.11.0

how it works

1. receiver
   • generates keys
   • posts its public key-set on server
     • public key pk, relinearization key rek, rotation key rok
2. sender
   • writes down some letter
   • in demo, random sample from enron dataset
   • tokenizes the letter
   • in demo, use already-tokenized value from enron dataset
   • make it one-or-zero(occur in letter or not) per embedding
   • gets public key from server
   • encrypt one-or-zero embedding
   • post result on server
3. receiver
   • get result from server
   • decrypt using secret key

how to run

NOTE that we made some update on ckks/params.go; see sender/setup.sh:17 and sender/params.go

server

cd your_path_to/HE_spamfilter/server

• install some golang dependency on main.go
• you need to handle some nasty PATH problem... see sender/setup.sh:11~16

cp ../sender/params.go your_GOPATH/src/github.com/ldsec/lattigo/v2/ckks/params.go

go run main.go

client - sender

cd your_path_to/HE_spamfilter/sender

• on linux, tar -zxvf enron.tar.gz
• on mac, gtar -zxvf enron.tar.gz
  • install GNU-tar using brew
  • brew update && brew install gnu-tar

docker run -it -v your_path_to/HE_spamfilter/sender:/home --rm ubuntu:focal

in container shell, cd /home bash setup.sh

client - receiver

docker run -it -v your_path_to/HE_spamfilter/receiver:/home --rm ubuntu:focal

in container shell, cd /home bash setup.sh

demo

with server & two client containers running,

• receiver python3 demo_keygen.py
• sender python3 demo_send.py
• receiver python3 demo_rcv.py

result

• accuracy
  • accuracy 0.96391
  • F1 score 0.96197
• latency
  • key post
    • public key (9.1 M) : 2.48 s
    • relinearization key (289 M) : 75 s
    • rotation key (289 M) : 52.87 s
    • by the way, secret key size is 4.6 M
  • key get
    • public key : 432.244 ms
  • message post
    • select random 10
    • POST + HE calculation + some filesystem access time
    • encrypted message : 5.1 M each
    • [1.878, 2.225, 1.636, 2.372, 1.838, 2.232, 1.860, 1.864, 1.839, 3.666] s
  • message get
    • calculation result : 5.1 M each
    • [167, 86, 103, 184, 81, 104, 299, 272, 96, 119] ms

about dataset

MWiechmann/enron_spam_data provides enron_spam_data.csv

We corrected some format error on that to make enron_spam_data_prep.csv, which consists of 33715 e-mails (16545 ham + 17170 spam)