We are a team from Wuhan University of technology, Wuhan, China.This project is mainly to solve the idash2021 competition - Track II: Homomorphic Encryption-based Secure Viral Strain Classification
| model\process | Preprocess [s] | Inference [s] |
|---|---|---|
| version1 | 14 | 0.172 |
| version2 | 100 | 8.2 |
| model | Accuracy plaintext&ciphtext |
|---|---|
| version1 | 98.6 & 98.55 |
| version2 | 99.5 & 99.05 |
put the hold-out dataset into the file folder.
docker build -t idash .
docker run -it idash /bin/bash
python3.8 preat.py #Get the test.txt file
g++ idash2021.cpp -p -main -ltfhe-spqlios-fma
./main
version1:
A strain containing 29000 base pairs is transformed into a 7500 dimensional vector as the input of the neural network.
B.1.427_5729�AGATCTGTTCTCTAAACGAACTTTAAAATCTGTGTGGCT.... （AAAA->1，AAAC->2,…,TTTT->256,other->0, If the dimension is less than 7500, supplement -1）
so,around(29000)x2000 --> 7500x2000
version2: If we are allowed to do excessive operations on preprocessing,we can also make statistics on the 7500 dimensional data obtained.
B.1.427_5729�AGATCTGTTCTCTAAACGAACTTTAAAATCTGTGTGGCT.... （AAAA->0，AAAC->1,…,TTTT->255，count the frequency of each gene string，each strain generated a 256 dimensional data）
around(29000)x2000 --> 256x2000
DNN model
On the ciphertext side, we use packaging technology，which greatly reduces the consumption of time.
If you choose to run our program on a physical machine,please install the tfhe library according to the following link and install python>=3.8.0. https://tfhe.github.io/tfhe/coding.html