HMV is an experimental hash-family extension of the Violet project. Where Violet explores a controller-driven cipher cascade, HMV carries the same design theme into a 1536-bit sponge permutation: an autonomous, coprime-period controller selects round constants and diffusion geometry.
Research software — do not deploy. HMV is a new construction, its security claims are design goals supported by the included analysis and formal artifacts, and it has not received independent cryptanalysis.
The HMV state is an 8 × 24 byte matrix (1536 bits) over GF(2^8). A
permutation applies sixteen rounds of:
A ← MixColumns(ShiftPlanes_t(SubBytes(A ⊕ K_t)))
MixColumns is an 8 × 8 circulant MDS matrix with branch number 9.
ShiftPlanes_t is a controller-selected row transposition, and K_t is a
controller-selected round constant. The controller combines a maximal-length
16-bit LFSR with a counter modulo 26, giving a period of 65,535 × 26 = 1,703,910 rounds. Its per-round offsets are constrained so that each input
column spreads across eight distinct output columns.
The repository contains a readable Python reference, a C implementation, test and known-answer tooling, analysis scripts, a paper, and a Lean development.
hmv/hmv.py Python reference implementation
src/ C implementation, self-test driver, and Makefile
analysis/ Parameter, diffusion, randomness, and generic experiments
figures/ Generated figures used below and by the paper
paper/ HMV paper and rendered PDF
lean/ Lean 4 formal development
The Python reference uses only the standard library.
import sys
sys.path.insert(0, "hmv")
from hmv import hmv, hmv_mac, hmvx
digest = hmv(b"abc", 256) # fixed-length digest
stream = hmvx(b"seed", 64) # 64-byte XOF output
tag = hmv_mac(b"key", b"message", 256)To build the C implementation, use a C11 compiler and make:
cd src
make
./hmv_test # inverse-permutation check, streaming check, KATs, benchmark
./hmv_test kat # known-answer output only
./hmv_test bench # benchmark onlymake also builds a constant-time configuration. The C and Python versions are
intended to agree on the included test vectors.
The figures below are regenerated by the scripts in analysis/. They describe
properties of this implementation and parameter set; they do not constitute a
security proof. See the paper for methods and the lean/
directory for formal artifacts.
Figure 1 — analysis/b0_parameters.py checks the selected S-box and diffusion
layer. The differential distribution table peaks at four, the Walsh spectrum
peaks at 32, and the search panel records the MDS candidates and their cost.
The script also exhaustively checks the selected matrix's square
subdeterminants.
Figure 2 — analysis/b1_diffusion.py follows the effect of a one-bit input
change. The first panel compares message and state diffusion by round, the
second visualizes which state bytes are reached, and the histogram checks the
per-output-bit flip probability against the strict-avalanche target of one half.
Figure 3 — analysis/b2_randomness.py runs a test battery over generated
streams. It compares observed pass counts with an expected baseline, displays
the p-value distribution, and plots byte-frequency deviations over the sampled
output. These are sanity checks, not evidence that the output is random.
Figure 4 — analysis/b3_generic.py compares observed truncated collisions
with birthday expectations, plots Hamming-weight changes caused by one flipped
message bit, and records implementation cost measurements. The collision and
avalanche results describe the sampled experiment rather than a full security
reduction.
Run from the HMV/ directory after installing the root project's Python
analysis dependencies:
python analysis/b0_parameters.py
python analysis/b1_diffusion.py
python analysis/b2_randomness.py
python analysis/b3_generic.pyThe scripts refresh the files in figures/ and summaries in analysis/results/.
They do not build or download Lean. If a Lean environment is already available,
its independent build metadata is in lean/.
Released under the Apache License 2.0.