Faster Homomorphic Trace-type Function Evaluation

This repository provides two sets of codes:

Implementation of faster trace (C++)

1-1) bench_unroll.cpp implementation of the unrolled trace

1-2) bench_seq.cpp implementation of the sequential trace (rotations-and-sums method)
Cost analysis in terms of # modular multiplications (python3)

1) Implementation with PALISADE (using the CKKS scheme)

Prerequisities

Modified version of PALSIADE v1.9.2

-> added faster automorphism via simple permutation (Permute function in poly.h/cpp and DCRTPoly.h).

The faster automorphism is applied to both 1-1) and 1-2) to conduct fair comparison.

Build and run

   mkdir build && cd build && cmake .. && make

Unrolling

  ./bench_unroll [logN] [M] [h] [L] [ell] [Delta] [d] [T]

e.g.

  ./bench_unroll 14 14 7 3 2 35 4 11

Rotations-and-sums

 ./bench_seq     [logN] [M]     [L] [ell] [Delta] [d] [T]

e.g.

 ./bench_seq 14 14 3 2 35 4 11

Description of each param

 log N:  ring dimension

 M: # iterations for rotations-and-sums

 h: # unrolled-iterations

 L: maximal level (# of RNS moduli for this level = L + 1)

 ell: level to examine (# of RNS moduli for this level = \ell + 1 where \ell <= L)

 Delta: scaling factor for the CKKS encoding

 d: maximal number of digits for key-switching (an integer that divides L+1)
 
 ---> # of special moduli k = (L+1)/d

 T: the number of experiments to perform

2) Cost Analysis

Reproducing Figures in Theoretical Analysis

  ./init_resulting_directory.sh

  cd ./cost_analysis

Figure 1 (Rate)

 python3 run_plot_fig1_ell_vs_rate.py

-> choose 1 and press enter

Plots are saved in /tmp/result/no_last_digit_rate/.

Figures 2 and 3 (Cost and Its Breakdown)

 python3 run_plot_fig2_fig3_dominant.py

Plots are saved in /tmp/result/cost_plot/ and /tmp/result/breakdown/.

Figure 4 (# Eval Keys)

 python3 run_plot_fig4_storage_various_M.py

Plots are saved in /tmp/result/numkey/.

uishi / FastHETrace