EPFL solution for iDash 2021 Task2

How to build the Docker

docker build -t epfl_solution .
docker run -it epfl_solution

Processing Samples

Samples are pre-processed by first applying a FCGR mapping, followed by a 2D DCTII [Lichtblau2019]. The top left h x h matrix of the DCTII (lowest frequencies) is extracted and set at the hash of the genome.

Training

$ go run model/main.go will process the samples of data/Challenge.fa and output the processed samples in model/X.binary model/Y.binary (X being the processed samples and Y the labels). The Python script model/training.py will use model/X.binary model/Y.binary that can then be used to train the model. The script will output the weights both in .npy and .binary as well as a .png image of the weights/features with gradient color coding.

Testing

$ make debug NBGENOMES=2000 will compile and run DebugTest.go which will process, encrypt, predict, decrypt the first 2000 samples located in data/Challenge.fa.

Run iDash21

$ make key : generates the secret-key and stores it in key/.
$ make pro NBGENOMES=2000 : processes the first 2000 samples located iin data/Challenge.fa. Returns the result in temps/.
$ make enc : Encrypts the processed samples. Returns the encrypted processed samples in temp/.
$ make pred : unmarshals the encrypted samples in temp/, evaluates the homomorphic prediction and marshals back the result in temp/.
$ make dec : unmarshals the encrypted prediction in temp/, decrypts and outputs the result in results/prediction.csv.

Parameters

Processing and crypto parameters are located in lib/params.

Others

$ make clean : clean all files in keys/, temps/,results/ and all compiled binary files. Does not clean files in model/.

Security

The HE evaluation security is based on the R-LWE hardness. The used parameters are log(N)=10, log(Q)=29. Both the secret and the Gaussian error are sampled from a truncated discrete Gaussian distribution with standard deviation 3.19 and bound 19. The security is estimated to 128-bit according to https://homomorphicencryption.org/.

References:

[Lichtblau2019] : Lichtblau Daniel. “Alignment-free genomic sequence comparison using FCGRand signal processing”, 2019.

Pro7ech/idash2021