diff --git a/README.md b/README.md index f745820..b2ea51b 100644 --- a/README.md +++ b/README.md @@ -32,7 +32,7 @@ | `nax-fl` 联邦聚合 + DP + FedAvg | ✅ 安全聚合责任链(Clip/Zero/SecureSum/Mean/GaussianNoise)+ 差分隐私(RDP 记账)+ 横向 FedAvg(local_train/metrics/加权均值/收敛轨迹),12 测含 4 逐位 golden;`--features secure` 另有 2 方 Vault 密态加权均值端到端 1 测 | | `nax-ml` 纵向 LR + MLP + 分箱 | ✅ sigmoid 奇多项式 + Beaver SGD(明文对照 >0.95);MLP 4 测;异构分箱 + WOE/IV + 互补抑制 6 测;密文 `Mᵀ@y` 箱计数 1 测 | | `nax-data` 数据层 | ✅ Frame(选择/矩阵/列和)+ 明文统计(均值/方差/协方差/相关)+ mock(确定性主键/schema 合成/DP)+ CSV,16 测 | -| `nax-sql` SCQL 引擎 | ✅ 手写解析器 + 掩码式 WHERE + 聚合/GROUP BY/k-匿名/投影约束/JOIN/HAVING/ORDER BY,15 测;`--features secure` 另有密态 WHERE+聚合 2 测 | +| `nax-sql` SCQL 引擎 | ✅ 手写解析器 + 掩码式 WHERE + 聚合/GROUP BY/k-匿名/投影约束/JOIN/HAVING/ORDER BY,15 测;`--features secure` 另有密态 WHERE+聚合+GROUP BY 3 测 | | `nax-policy` 策略/审计 | ✅ 出域守卫(exec/transfer/fetch)+ Label 六档 + 审计日志(过滤/汇总/裁剪),9 测 | | `nax-runtime` 执行模式 | ✅ dry_run / real_run(线程局部 + RAII 嵌套恢复),4 测 | | `nax-he` Paillier/OLE | ✅ 无可信发牌方 Beaver 物料 + PIR,50 条 Python 逐位对照 + 13 语义测试 | diff --git a/crates/nax-sql/src/secure.rs b/crates/nax-sql/src/secure.rs index 295bbc3..28cd8cd 100644 --- a/crates/nax-sql/src/secure.rs +++ b/crates/nax-sql/src/secure.rs @@ -10,6 +10,7 @@ //! SUM = Σ masked (本地线性求和,fxp) //! AVG = SUM / COUNT (COUNT 公开,scale_public + truncate) //! MIN/MAX = 对 masked 做两两 select_raw 归约,非命中行用 ±SENTINEL 补位 +//! GROUP BY = 逐公开码 c 的 bit_eq 掩码 AND 基础掩码,再走上面的聚合 //! ``` //! //! 只在 `feature = "secure"` 下编译,默认构建不引入 `nax-vault`。 @@ -41,7 +42,7 @@ pub enum Agg { /// MIN/MAX 补位哨兵(远大于常规业务值,又远小于定点溢出上限)。 const SENTINEL: f64 = (1u64 << 30) as f64; -/// 计算等值/比较掩码(共享原始 0/1 位)。 +/// 计算比较掩码(共享原始 0/1 位)。 pub fn mask_for( vault: &mut Vault, col: &[u64], @@ -67,29 +68,20 @@ pub fn mask_for( } } -/// 密态执行 `SELECT FROM t WHERE col T`(无 GROUP BY),返回各聚合的明文结果。 -/// -/// * `col`:本方的列分片(fxp 尺度); -/// * `thresh`:本方的阈值分片(fxp 尺度,长度与 `col` 相同); -/// * 结果按 `aggs` 顺序返回;AVG/MIN/MAX 在命中 0 行时返回 `NaN`。 -pub fn secure_where_aggregate( +/// 在给定 0/1 掩码下对 `col` 求各聚合,返回明文结果。 +fn aggregate_with_mask( vault: &mut Vault, col: &[u64], - thresh: &[u64], - op: CmpOp, + mask: &[u64], aggs: &[Agg], ) -> Result, VaultError> { let fxp = vault.fxp(); let len = col.len(); - if thresh.len() != len { - return Err(VaultError::Protocol("col 与 thresh 长度不一致".into())); - } - let mask = mask_for(vault, col, thresh, op)?; - let masked = vault.bit_mask(&mask, col)?; // mask · col,精确 + let masked = vault.bit_mask(mask, col)?; // mask · col,精确 // COUNT:本地线性求和(raw 整数分片)后 reveal let mut count_sh = 0u64; - for &m in &mask { + for &m in mask { count_sh = count_sh.wrapping_add(m); } let count = vault.reveal(&[count_sh])?[0] as usize; @@ -113,7 +105,6 @@ pub fn secure_where_aggregate( for &v in &masked { s = s.wrapping_add(v); } - // SUM × (1/count) → 2·fxp,再截回 fxp let scaled = vault.scale_public(&[s], 1.0 / count as f64)?; let tr = vault.truncate(&scaled)?; ring::decode(vault.reveal(&tr)?[0], fxp) @@ -124,7 +115,7 @@ pub fn secure_where_aggregate( f64::NAN } else { // 非命中行补 ±SENTINEL:(1−mask)·(±SENTINEL) 是本地线性 - let one_minus = vault.bit_not(&mask)?; + let one_minus = vault.bit_not(mask)?; let pad_sign = if agg == Agg::Min { SENTINEL } else { -SENTINEL }; let pad = vault.scale_public(&one_minus, pad_sign)?; let padded = vault.add_local(&masked, &pad)?; @@ -133,9 +124,9 @@ pub fn secure_where_aggregate( for i in 1..len { let cur = padded[i..i + 1].to_vec(); let pick_cur = if agg == Agg::Min { - vault.bit_lt(&cur, &best)? // cur < best + vault.bit_lt(&cur, &best)? } else { - vault.bit_lt(&best, &cur)? // best < cur + vault.bit_lt(&best, &cur)? }; best = vault.select_raw(&pick_cur, &cur, &best)?; } @@ -148,6 +139,56 @@ pub fn secure_where_aggregate( Ok(out) } +/// 密态执行 `SELECT FROM t WHERE col T`(无 GROUP BY),返回各聚合的明文结果。 +/// +/// `col` 是本方列分片(fxp 尺度);`threshold` 是公开字面量。 +pub fn secure_where_aggregate( + vault: &mut Vault, + col: &[u64], + threshold: f64, + op: CmpOp, + aggs: &[Agg], +) -> Result, VaultError> { + let th = vault.constant(threshold, col.len())?; + let mask = mask_for(vault, col, &th, op)?; + aggregate_with_mask(vault, col, &mask, aggs) +} + +/// 密态 GROUP BY:`gcol` 是已编码到 `0..n_groups-1` 的分组码分片(fxp 尺度), +/// `n_groups` 是**公开**取值域大小。`pred` 可选 `(threshold, op)` 作为 WHERE。 +/// +/// 返回 `n_groups` 行,每行按 `aggs` 顺序给出明文聚合值(组顺序 = 码顺序)。 +pub fn secure_group_by( + vault: &mut Vault, + col: &[u64], + gcol: &[u64], + n_groups: usize, + pred: Option<(f64, CmpOp)>, + aggs: &[Agg], +) -> Result>, VaultError> { + let len = col.len(); + if gcol.len() != len { + return Err(VaultError::Protocol("col 与 gcol 长度不一致".into())); + } + // 基础掩码:WHERE 或全 1 + let base = match pred { + Some((t, op)) => { + let th = vault.constant(t, len)?; + mask_for(vault, col, &th, op)? + } + None => vault.bit_one(len), + }; + + let mut out = Vec::with_capacity(n_groups); + for g in 0..n_groups { + let gc = vault.constant(g as f64, len)?; + let gmask = vault.bit_eq(gcol, &gc)?; + let mask = vault.bit_and(&base, &gmask)?; + out.push(aggregate_with_mask(vault, col, &mask, aggs)?); + } + Ok(out) +} + #[cfg(test)] mod tests { use super::*; @@ -155,14 +196,18 @@ mod tests { use nax_vault::Pipe; use std::thread; - /// 2 方 in-process:密态 `WHERE + 聚合` 必须等于明文 `Engine` 的结果。 + fn split_vec(va: &mut Vault, vals: &[f64], fxp: u32) -> (Vec, Vec) { + let enc: Vec = vals.iter().map(|v| ring::encode(*v, fxp).unwrap()).collect(); + let all = va.split(&enc).unwrap(); + let len = enc.len(); + (all[0..len].to_vec(), all[len..2 * len].to_vec()) + } + #[test] fn secure_where_aggregate_matches_engine() { let amounts = [100.0f64, 500.0, 250.0, 600.0, 300.0]; - let threshold = 400.0f64; let aggs = [Agg::Count, Agg::Sum, Agg::Avg, Agg::Min, Agg::Max]; - // 明文引擎的结果 let t = Table::new("t", "alice").with_col("amount", ColData::Num(amounts.to_vec())); let mut e = Engine::new(0); e.register(t); @@ -178,51 +223,33 @@ mod tests { ]; assert_eq!(want, [3.0, 650.0, 650.0 / 3.0, 100.0, 300.0]); - // 密态:2 方各自持分片 let (pa, pb) = Pipe::pair(); let sites = ["alice", "bob"]; let fxp = 18u32; let seed = [23u8; 32]; let mut va = Vault::new(pa, "alice", &sites, seed, fxp).unwrap(); let mut vb = Vault::new(pb, "bob", &sites, seed, fxp).unwrap(); - - let len = amounts.len(); - let enc: Vec = amounts.iter().map(|v| ring::encode(*v, fxp).unwrap()).collect(); - let enc_t = vec![ring::encode(threshold, fxp).unwrap(); len]; - let all = va.split(&enc).unwrap(); - let tall = va.split(&enc_t).unwrap(); - let c0 = all[0..len].to_vec(); - let c1 = all[len..2 * len].to_vec(); - let t0 = tall[0..len].to_vec(); - let t1 = tall[len..2 * len].to_vec(); + let (c0, c1) = split_vec(&mut va, &amounts, fxp); let ha = thread::spawn(move || { - secure_where_aggregate(&mut va, &c0, &t0, CmpOp::Lt, &aggs).unwrap() + secure_where_aggregate(&mut va, &c0, 400.0, CmpOp::Lt, &aggs).unwrap() }); let hb = thread::spawn(move || { - let _ = secure_where_aggregate(&mut vb, &c1, &t1, CmpOp::Lt, &aggs).unwrap(); + let _ = secure_where_aggregate(&mut vb, &c1, 400.0, CmpOp::Lt, &aggs).unwrap(); }); let got = ha.join().unwrap(); hb.join().unwrap(); - // COUNT/SUM/MIN/MAX 精确;AVG 有定点误差(scale_public 编码 1/3) assert_eq!(got[0], want[0], "COUNT"); assert!((got[1] - want[1]).abs() < 1e-6, "SUM {}", got[1]); - assert!( - (got[2] - want[2]).abs() < 1e-2, - "AVG {} vs {}", - got[2], - want[2] - ); + assert!((got[2] - want[2]).abs() < 1e-2, "AVG {} vs {}", got[2], want[2]); assert_eq!(got[3], want[3], "MIN"); assert_eq!(got[4], want[4], "MAX"); } - /// 无命中行:COUNT=0,SUM=0,AVG/MIN/MAX 为 NaN。 #[test] fn secure_where_empty_selection() { let vals = [10.0f64, 20.0, 30.0]; - let threshold = 5.0f64; // 全部 > 5 ⇒ col < 5 无命中 let aggs = [Agg::Count, Agg::Sum, Agg::Avg, Agg::Min, Agg::Max]; let (pa, pb) = Pipe::pair(); @@ -231,21 +258,14 @@ mod tests { let seed = [29u8; 32]; let mut va = Vault::new(pa, "alice", &sites, seed, fxp).unwrap(); let mut vb = Vault::new(pb, "bob", &sites, seed, fxp).unwrap(); - let len = vals.len(); - let enc: Vec = vals.iter().map(|v| ring::encode(*v, fxp).unwrap()).collect(); - let enc_t = vec![ring::encode(threshold, fxp).unwrap(); len]; - let all = va.split(&enc).unwrap(); - let tall = va.split(&enc_t).unwrap(); - let c0 = all[0..len].to_vec(); - let c1 = all[len..2 * len].to_vec(); - let t0 = tall[0..len].to_vec(); - let t1 = tall[len..2 * len].to_vec(); + let (c0, c1) = split_vec(&mut va, &vals, fxp); let ha = thread::spawn(move || { - secure_where_aggregate(&mut va, &c0, &t0, CmpOp::Lt, &aggs).unwrap() + // 全部 > 5 ⇒ col < 5 无命中 + secure_where_aggregate(&mut va, &c0, 5.0, CmpOp::Lt, &aggs).unwrap() }); let hb = thread::spawn(move || { - let _ = secure_where_aggregate(&mut vb, &c1, &t1, CmpOp::Lt, &aggs).unwrap(); + let _ = secure_where_aggregate(&mut vb, &c1, 5.0, CmpOp::Lt, &aggs).unwrap(); }); let got = ha.join().unwrap(); hb.join().unwrap(); @@ -256,4 +276,56 @@ mod tests { assert!(got[3].is_nan(), "MIN 应为 NaN"); assert!(got[4].is_nan(), "MAX 应为 NaN"); } + + #[test] + fn secure_group_by_matches_engine() { + let x = [10.0f64, 20.0, 30.0, 40.0, 50.0]; + let g = [0.0f64, 1.0, 0.0, 1.0, 0.0]; // 组码 + let aggs = [Agg::Count, Agg::Sum]; + + let t = Table::new("t", "alice") + .with_col("x", ColData::Num(x.to_vec())) + .with_col("g", ColData::Num(g.to_vec())) + .with_public("g"); + let mut e = Engine::new(0); + e.register(t); + let rt = e + .sql("SELECT g, COUNT(*) AS n, SUM(x) AS s FROM t GROUP BY g") + .unwrap(); + // g=0: rows 0,2,4 -> n=3, s=90;g=1: rows 1,3 -> n=2, s=60 + let want: Vec<(f64, f64)> = (0..rt.nrows()) + .map(|r| { + ( + rt.get(r, "n").unwrap().as_num().unwrap(), + rt.get(r, "s").unwrap().as_num().unwrap(), + ) + }) + .collect(); + assert_eq!(want, vec![(3.0, 90.0), (2.0, 60.0)]); + + let (pa, pb) = Pipe::pair(); + let sites = ["alice", "bob"]; + let fxp = 18u32; + let seed = [37u8; 32]; + let mut va = Vault::new(pa, "alice", &sites, seed, fxp).unwrap(); + let mut vb = Vault::new(pb, "bob", &sites, seed, fxp).unwrap(); + let (x0, x1) = split_vec(&mut va, &x, fxp); + let (g0, g1) = split_vec(&mut va, &g, fxp); + + let ha = thread::spawn(move || { + secure_group_by(&mut va, &x0, &g0, 2, None, &aggs).unwrap() + }); + let hb = thread::spawn(move || { + let _ = secure_group_by(&mut vb, &x1, &g1, 2, None, &aggs).unwrap(); + }); + let got = ha.join().unwrap(); + hb.join().unwrap(); + + assert_eq!(got.len(), 2); + for (gi, row) in got.iter().enumerate() { + let (wn, ws) = want[gi]; + assert_eq!(row[0], wn, "group {gi} COUNT"); + assert!((row[1] - ws).abs() < 1e-6, "group {gi} SUM {} vs {ws}", row[1]); + } + } } diff --git a/crates/nax-vault/src/compare.rs b/crates/nax-vault/src/compare.rs index 8dbe534..df31b56 100644 --- a/crates/nax-vault/src/compare.rs +++ b/crates/nax-vault/src/compare.rs @@ -41,6 +41,19 @@ impl Vault { } } + /// 公开常数在 fxp 尺度的分片(0 号持有编码值,其余为 0)。对应参考实现 `vault.constant`。 + /// + /// 公开常量共享成 `(c, 0, …)` 是安全的:重建恰好等于 `c`,且不新增任何隐私信息。 + pub fn constant(&self, value: f64, len: usize) -> Result, VaultError> { + let enc = nax_core::ring::encode(value, self.fxp) + .map_err(|e| VaultError::Crypto(e.to_string()))?; + if self.my_idx == 0 { + Ok(vec![enc; len]) + } else { + Ok(vec![0u64; len]) + } + } + /// 公开 0 的分片(全 0)。 pub fn bit_zero(&self, len: usize) -> Vec { vec![0u64; len] diff --git a/docs/GAP_PLAN.md b/docs/GAP_PLAN.md index 4f42b23..24a4725 100644 --- a/docs/GAP_PLAN.md +++ b/docs/GAP_PLAN.md @@ -92,7 +92,8 @@ Python 参考版 `nax/` 共 **8,963 行**。此前约 **4,150 行(≈46%)未 把明文引擎的核心语义搬到真实分片上,跑 2 方 in-process 会话对拍明文: - **`nax-sql` 加可选 `secure` feature**:`secure::secure_where_aggregate` 实现 - `SELECT FROM t WHERE col T`:掩码 = `bit_lt/gt/eq` → `bit_mask` → 本地线性求和/`scale_public`+`truncate`;MIN/MAX 用 `(1−mask)·(±SENTINEL)` 本地补位后两两 `bit_lt`+`select_raw` 归约。2 测对拍明文 `Engine`(含空命中)。 + `SELECT FROM t WHERE col T`:掩码 = `bit_lt/gt/eq` → `bit_mask` → 本地线性求和/`scale_public`+`truncate`;MIN/MAX 用 `(1−mask)·(±SENTINEL)` 本地补位后两两 `bit_lt`+`select_raw` 归约。 + `secure::secure_group_by`:逐**公开码** c 做 `bit_eq(gcol, c)` 与 WHERE 掩码 AND,再走同一聚合核心。阈值/组码用 `nax-vault::constant`(公开常量分片 `(c,0,…)`),无需调用方 split。3 测对拍明文 `Engine`(WHERE+聚合、空命中 NaN、GROUP BY)。 - **`nax-ml::binning_secure`**:`secure_bin_counts` 实现 `bad = Mᵀ·y`(`matmul_trunc`,转置为本地置换)+ `total`(本地列和)。1 测对拍明文 `Mᵀ@y` / `Mᵀ@1`。`nax-ml` 已硬依赖 `nax-vault`,无需 feature。 - 尺度约定:one-hot 与标签都按 fxp 编码(`1 → 2^fxp`),`Mᵀ@y` 落在 `2·fxp`,`matmul_trunc` 截回 `fxp`;`total` 线性列和天然 `fxp`。 -- 现状:`cargo test --workspace` 216 测;`-p nax-sql --features secure` 17 测、`-p nax-fl --features secure` 13 测、`-p nax-ml` 15 测,全绿、零警告。 +- 现状:`cargo test --workspace` 216 测;`-p nax-sql --features secure` 18 测、`-p nax-fl --features secure` 13 测、`-p nax-ml` 15 测,全绿、零警告。