全同态

文件	最后提交记录	最后更新时间
cmake	HiFHE-bgv-migration fix problems	23 天前
include	HiFHE-bgv-migration fix problems	23 天前
src	fix(bgv): final_report findings, plus a benchmark harness Review fixes ------------ Memory safety and correctness: - The vector constructor sized its allocation in uint32 while the fill counted in size_t, so a count past UINT32_MAX/sizeof(T) wrapped the allocation and the fill ran off it. ASan reports the overflow without the bound. - Relinearize built its output through a constructor that zeroes caps and never restored them, so a slot-mapped ciphertext came out marked plain and decoded in the wrong order. - Decryption multiplies both operands as NTT values but accepted coefficient-form input, returning nonsense with success. - The relin gadget is built for one chain length; applying it to any size-3 product made multiplying mod-switched ciphertexts fail for holders of a relin key and work for everyone else. - A duplicated context lost its seed, so the same seed stopped giving the same keys. - Decryption asks for every secret-key power up to a wire-supplied component count, and each cost a multiplication per step: quadratic in the input. Square-and-multiply instead. Secret hygiene, matching what the rest of the tree does: plaintext storage (secret keys and decrypted messages both) and the secret-derived polynomials in keygen are wiped before release. Views are left alone, since their storage belongs to the key itself. Smaller items: a duplicate shares the cache handle it was given; the fixed-weight sampler enforces the same label and seed-length contract as the other samplers; OCTETS param buffers are read without assuming alignment; a lifecycle case no longer describes keygen as an M0 stub. Six regression tests, each shown to fail first. Full SDV 67/67 under ASan. Benchmark --------- testcode/benchmark, modelled on platform/openhitls/testcode/benchmark: a module registers setUp/tearDown and a list of operations, the driver walks the registry and prints one CSV row per measurement. - Kernel: the transforms and modular arithmetic under them (NTT forward and inverse, table build, MulDot, Add), with serialization out of the picture. - Bgv: what a caller going through CRYPT_BGV_* pays, wire encoding included. Driven by testcode/script/build_benchmark.sh and execute_benchmark.sh. It also builds inside the testcode tree with -DBUILD_BENCHMARK=ON, and forces -O2 on itself either way, since a benchmark at -O0 measures the compiler. RLWE_RnsPoly_Create and _Destroy come back for it: the kernel module drives one modulus directly, which is what they are for. They were removed as unused when this consumer had not been ported yet. The benchmark honours the same -DHIFHE_<ALG> switches the provider has. A module calls its algorithm directly, so it is compiled and registered only when that algorithm is in the provider: both switches default to what hifhe_config.h says and can be forced either way, as CMake options (-DHIFHE_BGV=OFF) or through the script (--disable bgv), matching build_hifhe.sh. Selecting nothing is refused at configure time rather than producing a binary with nothing to measure. Without this a provider built --disable bgv left the benchmark failing to link, the same way the SDV suites used to.	18 天前
testcode	fix(bgv): final_report findings, plus a benchmark harness Review fixes ------------ Memory safety and correctness: - The vector constructor sized its allocation in uint32 while the fill counted in size_t, so a count past UINT32_MAX/sizeof(T) wrapped the allocation and the fill ran off it. ASan reports the overflow without the bound. - Relinearize built its output through a constructor that zeroes caps and never restored them, so a slot-mapped ciphertext came out marked plain and decoded in the wrong order. - Decryption multiplies both operands as NTT values but accepted coefficient-form input, returning nonsense with success. - The relin gadget is built for one chain length; applying it to any size-3 product made multiplying mod-switched ciphertexts fail for holders of a relin key and work for everyone else. - A duplicated context lost its seed, so the same seed stopped giving the same keys. - Decryption asks for every secret-key power up to a wire-supplied component count, and each cost a multiplication per step: quadratic in the input. Square-and-multiply instead. Secret hygiene, matching what the rest of the tree does: plaintext storage (secret keys and decrypted messages both) and the secret-derived polynomials in keygen are wiped before release. Views are left alone, since their storage belongs to the key itself. Smaller items: a duplicate shares the cache handle it was given; the fixed-weight sampler enforces the same label and seed-length contract as the other samplers; OCTETS param buffers are read without assuming alignment; a lifecycle case no longer describes keygen as an M0 stub. Six regression tests, each shown to fail first. Full SDV 67/67 under ASan. Benchmark --------- testcode/benchmark, modelled on platform/openhitls/testcode/benchmark: a module registers setUp/tearDown and a list of operations, the driver walks the registry and prints one CSV row per measurement. - Kernel: the transforms and modular arithmetic under them (NTT forward and inverse, table build, MulDot, Add), with serialization out of the picture. - Bgv: what a caller going through CRYPT_BGV_* pays, wire encoding included. Driven by testcode/script/build_benchmark.sh and execute_benchmark.sh. It also builds inside the testcode tree with -DBUILD_BENCHMARK=ON, and forces -O2 on itself either way, since a benchmark at -O0 measures the compiler. RLWE_RnsPoly_Create and _Destroy come back for it: the kernel module drives one modulus directly, which is what they are for. They were removed as unused when this consumer had not been ported yet. The benchmark honours the same -DHIFHE_<ALG> switches the provider has. A module calls its algorithm directly, so it is compiled and registered only when that algorithm is in the provider: both switches default to what hifhe_config.h says and can be forced either way, as CMake options (-DHIFHE_BGV=OFF) or through the script (--disable bgv), matching build_hifhe.sh. Selecting nothing is refused at configure time rather than producing a binary with nothing to measure. Without this a provider built --disable bgv left the benchmark failing to link, the same way the SDV suites used to.	18 天前
.clang-format	init repo	1 个月前
.gitignore	init repo	1 个月前
CMakeLists.txt	HiFHE-bgv-migration fix problems	23 天前
LICENSE	Initial commit	1 个月前
README-zh.md	docs: Add BGV capability description Co-authored-by: zyw007<zhaoyiwei4@h-partners.com>	20 天前
README.md	docs: Add BGV capability description Co-authored-by: zyw007<zhaoyiwei4@h-partners.com>	20 天前
SECURITY.md	Add security policy	25 天前
SECURITY_zh.md	Add security policy	25 天前

HiFHE

HiFHE 是面向 openHiTLS 的同态加密 provider 密码库。目前提供 CKKS 和 BGV 算法。

功能

1. 公共能力

对接 openHiTLS provider，提供算法注册、查询及 provider 生命周期管理。
提供已编译算法能力查询接口 HIFHE_GetFeatures。
支持动态库、静态库、算法裁剪、ASan 和覆盖率构建。
提供基于 openHiTLS SDV 框架的生成式测试。

2. 当前支持的算法与能力

CKKS

参数配置、算法上下文创建、复制和释放。
密钥生成、公私钥导入导出。
复数向量编码与解码、公钥加密和私钥解密。
通过 openHiTLS provider 调用密钥管理和加解密接口。

尚未提供密文同态加法、乘法、旋转、重缩放及自举接口。

BGV

参数配置、算法上下文创建、复制和释放。
密钥生成、公私钥导入导出。
整数向量的批编码与解码、公钥加密和私钥解密。
密文加乘、重线性化、模切换、旋转和重随机化。

构建

依赖 CMake 3.16 或更高版本、C11 编译器和 openHiTLS。不传 --hitls_dir 时，脚本会下载并编译 openHiTLS。默认同时生成 libhifhe_provider.so 和 libhifhe_provider.a。

bash testcode/script/build_hifhe.sh

使用已有 openHiTLS：

bash testcode/script/build_hifhe.sh --hitls_dir /path/to/openHiTLS
# 当前工作区：
bash testcode/script/build_hifhe.sh --hitls_dir /usr1/cq/openhitls/openhitls

只生成一种库：

bash testcode/script/build_hifhe.sh shared
bash testcode/script/build_hifhe.sh static

使用已编译的 openHiTLS 直接执行 CMake：

cmake -S . -B build \
    -DHITLS_ROOT_PATH=/path/to/openHiTLS \
    -DHITLS_LIB_DIR=/path/to/openHiTLS/build \
    -DHIFHE_BUILD_SHARED=ON \
    -DHIFHE_BUILD_STATIC=ON \
    -DHIFHE_CKKS=ON \
    -DHIFHE_BGV=ON \
    -DHIFHE_ENABLE_ASAN=OFF \
    -DHIFHE_ENABLE_GCOV=OFF
cmake --build build --parallel

CMake 配置：

变量	默认值	说明
`HITLS_ROOT_PATH`	`platform/openhitls`	openHiTLS 源码目录
`HITLS_LIB_DIR`	`${HITLS_ROOT_PATH}/build`	`hitls_crypto`、`hitls_bsl` 库目录
`HIFHE_BUILD_SHARED`	`ON`	生成 `libhifhe_provider.so`
`HIFHE_BUILD_STATIC`	`ON`	生成 `libhifhe_provider.a`
`HIFHE_CKKS`	`ON`	将 CKKS 编译进 provider
`HIFHE_BGV`	`ON`	将 BGV 编译进 provider
`HIFHE_ENABLE_ASAN`	`OFF`	启用 AddressSanitizer
`HIFHE_ENABLE_GCOV`	`OFF`	启用覆盖率

至少启用一种库类型。CKKS 和 BGV 均支持按算法裁剪：

bash testcode/script/build_hifhe.sh --enable ckks
bash testcode/script/build_hifhe.sh --disable ckks
bash testcode/script/build_hifhe.sh --enable bgv
bash testcode/script/build_hifhe.sh --disable bgv

HIFHE_CKKS=OFF 或 HIFHE_BGV=OFF 会排除对应算法源码，并将结果写入生成的 build/include/hifhe_config.h。

安装命令：cmake --install build --prefix /path/to/prefix。

测试

编译：

bash testcode/script/build_sdv.sh

用例依赖动态 provider。只编译一个套件：

bash testcode/script/build_sdv.sh run-tests=test_suite_sdv_hifhe_ckks

运行：

bash testcode/script/execute_sdv.sh

项目介绍

全同态

MulanPSL-2.0 C 16提交数bgv

定制我的领域

README

规则集

HiFHE:基于 C 密码生态的同态加密 provider 项目

HiFHE

功能

1. 公共能力

2. 当前支持的算法与能力

CKKS

BGV

构建

测试

项目介绍