# Eight Implementations, One Cipher: A Cross-Language Differential Study of SM4 (GB/T 32907-2016) Author: Daozhendong Yang (yxpil), Yunnan Technology and Business University — hi@yxpil.com Artifact: https://github.com/yxpil/SM4Kit ## Suggested arXiv classification - **Primary:** `cs.CR` (Cryptography and Security) - **Cross-list:** `cs.SE` (Software Engineering) - **Cross-list:** `cs.PL` (Programming Languages) The same categories are recorded in the header comment of `main.tex`. ## Abstract (verbatim from `main.tex`) SM4 is a 128-bit block cipher standardised in GB/T 32907-2016. SM4Kit is an open-source library that implements it eight times --- in JavaScript, Python, Go, Java, Rust, PHP, Ruby and C --- behind one API, with no third-party dependencies. This paper is not a proposal for a cipher and claims no cryptographic novelty; it is a differential-testing study of those eight implementations. We describe a harness that drives each of them through the GB/T 32907 Annex A vectors, a deterministic corpus of 256 random keys and blocks, and nine ECB/CBC payloads at boundary lengths, and that compares ciphertexts, the full 32-word round-key schedule, and mode outputs against one another and against a harness-local reference model. Four toolchains (JavaScript, Python, Java, Rust) were available on the authoring machine and were executed; Go, PHP, Ruby and C were read and statically checked but *not executed*, and are reported as not executed rather than as passes. For the four executed implementations the harness finds no divergence at all: 259 block cases, 8,288 round-key words and 18 mode outputs per language agree exactly, and all six language pairs agree pairwise. A static extraction of the S-box table from all eight sources finds a single 256-entry bijection, identical everywhere. To show that this null result is not vacuous we inject 294 single-point mutants into the reference model: the standard vector alone detects 244 of them (83.0%), missing 50 of the 256 S-box table mutations, while a 16-case differential corpus detects 294 of 294. The harness, corpus, drivers, raw transcripts and figure scripts are released so that a reader with the four missing toolchains can complete the matrix. Note: in the LaTeX source `---` is an em dash, `*not executed*` is `\emph{not executed}`, and `8,288` is `8{,}288`. Wording is otherwise identical. ## File manifest ``` papers/sm4kit/ main.tex the paper (two-column, article, 10pt) refs.bib 16 bibliography entries (GB/T 32907, differential testing, crypto validation, modes/padding) common.tex shared preamble/macros, copied from papers/_shared/common.tex README.md this file figures/ loc.pdf Figure 1: core vs self-test lines of code per language matrix.pdf Figure 2: executed/pass/not-run matrix, 8 languages x 9 checks mutation.pdf Figure 3: S-box mutant detection vs corpus size scripts/ reference.py harness-local reference model of SM4 (ninth oracle), S-box transcribed from GB/T 32907 Table 6 differential.py the single harness: builds the corpus, runs toolchains, drives per-language drivers, compares ciphertexts, round keys and mode outputs, runs static checks and the mutation study, writes data/*.json and data/*.csv make_figures.py regenerates the three PDF figures from data/*.json drivers/ driver.js JavaScript driver (node) driver.go Go driver SM4Driver.java Java driver driver.rs Rust driver driver.php PHP driver driver.rb Ruby driver driver.c C driver (Python has no driver process: it is imported and driven in-process by differential.py, because the interpreter is already running.) data/ differential.json full results: status, per-language results, pairwise agreement, S-box hashes, LOC, self-test outcomes differential.csv the same per-case results in flat form mutation.json the 294-mutant detection study jobs/cases.tsv the exact block corpus driven (key, block, expected) jobs/payloads.tsv the exact ECB/CBC payload corpus driven transcripts/*.tsv raw driver output, one file per executed language logs/*-selftest.log raw self-test output, one file per executed language harness-run.log console transcript of the final harness run ``` ## Build The paper needs only a standard TeX Live / MiKTeX installation; it uses `abbrvnat`, `natbib`, `amsmath`/`amssymb`, `graphicx`, `booktabs`, `url`, `hyperref` and `enumitem` (all in `common.tex` plus one extra package). ```bash cd papers/sm4kit pdflatex main.tex bibtex main pdflatex main.tex pdflatex main.tex ``` The build was **not** executed as part of preparing this package: no LaTeX compiler was installed on the authoring machine. The source has therefore been checked by reading, not by compiling; if a warning or error appears, it is the first thing to fix. ## Packaging for submission From the repository root: ```bash bash papers/make_package.sh sm4kit ``` ## Submission Upload the generated archive at https://arxiv.org/submit (choose `cs.CR` as primary, `cs.SE` and `cs.PL` as cross-lists). arXiv compiles the source itself; `main.tex`, `refs.bib`, `common.tex` and `figures/*.pdf` must all be present in the archive. The abstract is 1,698 characters, inside arXiv's 1,920-character form limit, so it can be pasted into the submission form unchanged. ## Reproduction Requirements: Python 3, plus whichever of `node`, `python`, `go`, `javac`/`java`, `rustc`, `php`, `ruby`, `gcc`/`cc`/`clang` are installed. Missing toolchains are detected and reported as `not_executed`; they are never reported as passes. ```bash cd papers/sm4kit # 1. Run the differential harness (writes data/differential.json, # data/differential.csv, data/mutation.json and the raw transcripts). python scripts/differential.py # 2. Regenerate the figures from those results. python scripts/make_figures.py # 3. Optional: repeat the differential corpus without re-running the # repository self-tests (they dominate the wall clock; the Python suite # alone takes about 48 s because of the 1,000,000-iteration vector). python scripts/differential.py --skip-selftest ``` Useful flags for `differential.py`: `--repo PATH` (default `_repos/SM4Kit`), `--cases N` (default 256 random keys/blocks), `--seed N` (default `0x534D34`). Status of the shipped run: JavaScript (node 24.21.0), Python (CPython 3.13.14), Java (JDK 27) and Rust (rustc 1.98.1) executed and passed; Go, PHP, Ruby and C were read and statically checked only, because no Go, PHP, Ruby or C toolchain was installed. To complete the matrix, install those four toolchains and re-run step 1; the drivers are already in `scripts/drivers/`. ## What this paper does not claim - **No new cipher and no cryptographic novelty.** SM4 is a published standard (GB/T 32907-2016). Nothing here improves, attacks or re-parameterises it. The contribution is the cross-language conformance methodology and the artefacts. - **No security claim about SM4Kit as a system.** The library provides ECB and CBC with PKCS#7 padding and no authentication; CBC without a MAC is exposed to padding-oracle attacks. The paper does not evaluate the library's security beyond conformance. - **No constant-time claim.** No timing, cache or side-channel measurement was performed; the implementations are ordinary table-driven code with a 256-entry S-box and are not claimed to be constant time. - **No claim that all eight languages passed.** Only four were executed. Go, PHP, Ruby and C are reported as not executed; their source was read and their S-box tables were extracted and compared, which is not execution. - **No claim that agreement proves correctness.** Agreement among implementations is evidence of conformance, not proof. A shared mistake inherited from the standard vector would be invisible to the pairwise comparison; the mutation study measures this blind spot (the standard vector alone misses 50 of 256 S-box mutants) rather than removing it. - **No claim of independence between the implementations.** They share one author, one reference and one review process; "separate" is intended throughout, not "independently developed" in the N-version sense. - **No performance or memory evaluation.** Timings quoted are wall-clock measurements of the harness on one Windows 11 x86-64 machine, not benchmarks. - **No claim about TypeScript.** The repository has a `.d.ts` declaration file only; there is no separate TypeScript implementation, so the abstract's list of eight languages is JavaScript, Python, Go, Java, Rust, PHP, Ruby and C.