fix(demo): 反面案例改 Rust、推理过程左对齐、首页清 Python 残留

三个用户反馈的修复,外加顺带挖出的三个真实缺陷。

反面案例默认示例(Python → Rust)
- 平台是 Rust 框架,审的代码绝大多数是 Rust;拿 Python 当默认样例会让人
  误以为「平台审的是 Python」,也让规则的 Rust 分支失去演示意义
- 新反例覆盖 F01–F11 全部规则,且故意用错 API 签名(blind 传 4 参、
  match_across 传 4 参、Handle::new 少 Label)—— 模型生成最常见的翻车点
- 线上实测:命中 F01–F11 全部 11 条,score 0 / 禁止入域

推理过程居中 → 左对齐
- 根因:thinking 的 span 继承了外层「推理中」的 text-align:center,
  一段英文推理被折成居中几行,左右大片空白,极难读
- 抽出 .thinking-panel 到 _shared/demo.css:显式 left + pre-wrap + 独立边框,
  像日志一样从左往右顺读;轮询刷新后自动滚到底部(tail 是尾部窗口)

首页 Python 残留
- 删 footer 的「本地启动:python demo/server.py」两行
- badge 的「PID」「Python 3.13.12」是前端硬编码假数据(后端早就是 Rust
  单二进制,既无 PID 概念也无 Python 运行时)→ 改读 /api/health 真值
- 修 refs/refs/refs/ 三重前缀 bug(真实路径 refs/nax/examples/)12 处
- 修编造的 crate 名:mpc-device→mpc-vault、mpc-privacy→mpc-fl、
  mpc-crypto→mpc-he、mpc-ml metrics/aggregators→mpc-fl、
  SecureSQL.explain→Engine::explain
- 「14 条静态 + 2 条 AST」是错的:规则引擎是 F01–F11,AST 规则根本不存在
  (14 是红线条数)→ 改为「F01–F11 静态规则 + 14 条红线 + LLM 语义复核」

顺带修掉的规则引擎盲区
- F04 硬编码密钥只认 Python 的 `api_key = "..."`,Rust 最常见的
  `const API_KEY: &str = "..."` 整个漏掉(等号前夹着类型标注就匹配不上)
  → 新增 skip_type_annotation(),兼容 Python / Rust const / Rust let 三种写法
- 补 f04_detects_rust_const_with_type_annotation 回归测试钉死

测试 35 passed / 0 failed;ci.sh 四条全绿;镜像 mpc-demo-v0.1.9 已上线
This commit is contained in:
admin committed 2026-11-06 14:02:33 +08:00
1 parent 6735b6c732
commit bc0a271cc4
6 files changed
+301 -84

No files matched your search

+51
View File
@@ -138,3 +138,54 @@ table.psi td.alice { color: var(--alice); font-weight: 600; }
.flow-sub { fill: var(--text-faint); font: 10px var(--mono); }
.l1-card-grid { display: grid; grid-template-columns: 1fr 1fr; gap: 16px; }
/* 大模型思考过程面板
*
* 为什么必须单独一个类:思考片段是**长文本**,而外层「推理中…」的提示
* 用的是 `text-align:center`。span 继承父级的居中后,一段英文推理会被
* 折成居中的几行,左右两侧全是空白,读起来非常难受 —— 模型吐的是
* 逐字推理,本来就该像日志一样从左往右顺着读。
*
* 所以这里显式 `text-align:left`,并给它独立的边框/底色,
* 让它一眼就是「另一块东西」,而不是提示文案的一部分。
*/
.thinking-panel {
margin-top: 10px;
background: var(--bg);
border: 1px solid var(--border);
border-radius: 6px;
text-align: left; /* 覆盖父级的居中 —— 这是本类的关键 */
overflow: hidden;
}
.thinking-panel .tp-head {
display: flex; align-items: center; gap: 8px;
padding: 5px 10px;
border-bottom: 1px solid var(--border);
background: var(--panel-2);
color: var(--text-faint);
font-family: var(--mono);
font-size: 10px;
letter-spacing: .4px;
}
.thinking-panel .tp-head .tp-count { margin-left: auto; }
.thinking-panel .tp-body {
padding: 9px 11px;
font-family: var(--mono);
font-size: 10.5px;
line-height: 1.65;
color: var(--text-dim);
white-space: pre-wrap; /* 保留模型推理的换行,不压成一行 */
word-break: break-word;
max-height: 200px;
overflow-y: auto;
}
.thinking-panel .tp-body:empty::before {
content: "(等待模型输出思考片段…)";
color: var(--text-faint);
}
/* 深色底上的滚动条别太扎眼 */
.thinking-panel .tp-body::-webkit-scrollbar { width: 8px; }
.thinking-panel .tp-body::-webkit-scrollbar-thumb {
background: var(--border); border-radius: 4px;
}
.thinking-panel .tp-body::-webkit-scrollbar-track { background: transparent; }
+38 -41
View File
@@ -167,8 +167,8 @@
</div>
<div class="badges">
<span class="badge"><span class="dot" id="status-dot"></span><span id="status-text">后端在线</span></span>
<span class="badge">PID <span id="status-pid">—</span></span>
<span class="badge">Python <span id="status-py">—</span></span>
<span class="badge" id="status-backend">后端 <span id="status-backend-name">—</span></span>
<span class="badge" id="status-secure">secure <span id="status-secure-val">—</span></span>
</div>
</header>
@@ -183,7 +183,7 @@
<a class="card" data-level="L3" href="/hello/">
<div class="head"><span class="num">01</span><span class="level">L3</span></div>
<h3><span class="ico">🔰</span>Hello MPC</h3>
<div class="name-en">refs/refs/refs/nax/examples/01_hello_mpc.py</div>
<div class="name-en">refs/nax/examples/01_hello_mpc.py</div>
<div class="desc">三方密文求和入门:<code>site.put</code> / <code>vault.share</code> / 密文四则运算 / <code>reveal</code>。</div>
<div class="mpc-ref">mpc 模块: <code>mpc.{init,vault,site}</code></div>
<span class="arrow">→</span>
@@ -191,7 +191,7 @@
<a class="card" data-level="L1" href="/psi/">
<div class="head"><span class="num">02</span><span class="level">L1</span></div>
<h3><span class="ico">🔐</span>PSI 求交</h3>
<div class="name-en">refs/refs/refs/nax/examples/02_psi_join.py</div>
<div class="name-en">refs/nax/examples/02_psi_join.py</div>
<div class="desc">ECDH-PSI 双方求交:bluff / reblind / match / resolve,7 阶段动画展示每行 ID 流转。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-psi</code></div>
<span class="arrow">→</span>
@@ -199,7 +199,7 @@
<a class="card" data-level="L4" href="/k8s_two_party/">
<div class="head"><span class="num">12</span><span class="level">L4</span></div>
<h3><span class="ico">⎈</span>k8s 跨方</h3>
<div class="name-en">refs/refs/refs/nax/examples/12_k8s_two_party</div>
<div class="name-en">refs/nax/examples/12_k8s_two_party</div>
<div class="desc">Kubernetes 跨方部署示例:两方数据分别部署在独立 namespace,ray cluster 调度 PSI 任务。</div>
<div class="mpc-ref">指引: <code>examples/12_k8s_two_party/README.md</code></div>
<span class="arrow">→</span>
@@ -207,7 +207,7 @@
<a class="card" data-level="L3" href="/dry_run/">
<div class="head"><span class="num">11</span><span class="level">L3</span></div>
<h3><span class="ico">🩺</span>Dry Run 诊断</h3>
<div class="name-en">refs/refs/refs/nax/examples/11_dry_run.py</div>
<div class="name-en">refs/nax/examples/11_dry_run.py</div>
<div class="desc">算法干运行:检查 cluster / backend / party connectivity,给出运行前诊断报告。</div>
<div class="mpc-ref">mpc 模块: <code>mpc dryrun</code></div>
<span class="arrow">→</span>
@@ -223,7 +223,7 @@
<a class="card" data-level="L1" href="/vertical_lr/">
<div class="head"><span class="num">03</span><span class="level">L1</span></div>
<h3><span class="ico">🏥</span>纵向 LR</h3>
<div class="name-en">refs/refs/refs/nax/examples/03_vertical_lr.py</div>
<div class="name-en">refs/nax/examples/03_vertical_lr.py</div>
<div class="desc">同一批用户特征分散在医院 + 保险,标签只在保险方,全程密文训练 + 明文对照组。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-ml VerticalLR</code></div>
<span class="arrow">→</span>
@@ -231,7 +231,7 @@
<a class="card" data-level="L2" href="/horizontal_fedavg/">
<div class="head"><span class="num">04</span><span class="level">L2</span></div>
<h3><span class="ico">🏦</span>横向 FedAvg</h3>
<div class="name-en">refs/refs/refs/nax/examples/04_horizontal_fedavg.py</div>
<div class="name-en">refs/nax/examples/04_horizontal_fedavg.py</div>
<div class="desc">三家分行客户不重叠、特征口径一致;对比 mask / paillier 两种安全聚合方式。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-fl FedAvg</code></div>
<span class="arrow">→</span>
@@ -239,7 +239,7 @@
<a class="card" data-level="L2" href="/dp_fedavg/">
<div class="head"><span class="num">06</span><span class="level">L2</span></div>
<h3><span class="ico">🛡</span>DP-FedAvg</h3>
<div class="name-en">refs/refs/refs/nax/examples/06_dp_fedavg.py</div>
<div class="name-en">refs/nax/examples/06_dp_fedavg.py</div>
<div class="desc">FedAvg + 差分隐私噪声;展示 ε-DP 预算与模型准确率的 trade-off 曲线。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-fl DPFedAvg</code></div>
<span class="arrow">→</span>
@@ -247,7 +247,7 @@
<a class="card" data-level="L2" href="/secure_mlp/">
<div class="head"><span class="num">07</span><span class="level">L2</span></div>
<h3><span class="ico">🧠</span>Secure MLP</h3>
<div class="name-en">refs/refs/refs/nax/examples/07_secure_mlp.py</div>
<div class="name-en">refs/nax/examples/07_secure_mlp.py</div>
<div class="desc">安全多方神经网络:每层密文推理、ReLU 多项式逼近,无中间明文出域。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-ml SecureMLP</code></div>
<span class="arrow">→</span>
@@ -263,15 +263,15 @@
<a class="card" data-level="L1" href="/pir/">
<div class="head"><span class="num">08</span><span class="level">L1</span></div>
<h3><span class="ico">🔍</span>PIR 隐匿查询</h3>
<div class="name-en">refs/refs/refs/nax/examples/08_pir.py</div>
<div class="name-en">refs/nax/examples/08_pir.py</div>
<div class="desc">1-out-of-N 隐匿查询:客户端 Paillier 单位向量加密,服务端密文上加权求和,服务端不知道查了谁。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-crypto pir</code></div>
<div class="mpc-ref">mpc 模块: <code>mpc-he pir</code></div>
<span class="arrow">→</span>
</a>
<a class="card" data-level="L1" href="/scql/">
<div class="head"><span class="num">09</span><span class="level">L1</span></div>
<h3><span class="ico">📊</span>SCQL 安全 SQL</h3>
<div class="name-en">refs/refs/refs/nax/examples/09_scql.py</div>
<div class="name-en">refs/nax/examples/09_scql.py</div>
<div class="desc">PSI-JOIN + WHERE-as-weight + GROUP BY;展示"在两方私有数据上直接写 SQL"的实际效果。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-sql scql</code></div>
<span class="arrow">→</span>
@@ -279,7 +279,7 @@
<a class="card" data-level="L2" href="/hetero_binning/">
<div class="head"><span class="num">10</span><span class="level">L2</span></div>
<h3><span class="ico">📐</span>异构分箱</h3>
<div class="name-en">refs/refs/refs/nax/examples/10_hetero_binning.py</div>
<div class="name-en">refs/nax/examples/10_hetero_binning.py</div>
<div class="desc">跨方分箱对齐:各自特征分箱不同,安全求交分箱边界 + IV 计算。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-ml HeteroBinning</code></div>
<span class="arrow">→</span>
@@ -287,7 +287,7 @@
<a class="card" data-level="L2" href="/policy_audit/">
<div class="head"><span class="num">05</span><span class="level">L2</span></div>
<h3><span class="ico">📋</span>策略审计</h3>
<div class="name-en">refs/refs/refs/nax/examples/05_policy_audit.py</div>
<div class="name-en">refs/nax/examples/05_policy_audit.py</div>
<div class="desc">数据访问策略审计:哪些字段、什么用途、什么角色、是否授权;输出审计报告。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-policy</code></div>
<span class="arrow">→</span>
@@ -305,15 +305,15 @@
<a class="card" data-level="L2" href="/secure_compare/">
<div class="head"><span class="num">13</span><span class="level">L2</span></div>
<h3><span class="ico">🔍</span>密文比较</h3>
<div class="name-en">mpc-device vault lt/gt/eq/select/max/argmax</div>
<div class="name-en">mpc-vault compare {bit_lt, bit_gt, bit_eq, msb_raw}</div>
<div class="desc">edaBit + BitLT 协议:PSI 共同客户中"谁存款最高",零泄露找出极值与下标。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-device vault {lt,gt,eq,select,amax,argmax}</code></div>
<div class="mpc-ref">mpc 模块: <code>mpc-vault compare {bit_lt,bit_gt,bit_eq,msb_raw}</code></div>
<span class="arrow">→</span>
</a>
<a class="card" data-level="L2" href="/secure_stats/">
<div class="head"><span class="num">14</span><span class="level">L2</span></div>
<h3><span class="ico">📊</span>安全统计</h3>
<div class="name-en">mpc-data stats {mean, var, pearson, correlation_matrix}</div>
<div class="name-en">mpc-data stats {mean, var_pop, pearson, correlation_matrix}</div>
<div class="desc">两方数据"年龄 vs 保费"相关性,全程密文计算;只 reveal 1 个皮尔逊 r。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-data stats</code></div>
<span class="arrow">→</span>
@@ -321,17 +321,17 @@
<a class="card" data-level="L2" href="/secure_metrics/">
<div class="head"><span class="num">15</span><span class="level">L2</span></div>
<h3><span class="ico">📈</span>密文模型评估</h3>
<div class="name-en">mpc-ml metrics.{accuracy, auc, log_loss, mse}</div>
<div class="name-en">mpc-fl metrics.{accuracy, auc, log_loss, mse}</div>
<div class="desc">联合训练完模型,看 ACC / AUC / log_loss / MSE 不暴露原始数据。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-ml metrics</code></div>
<div class="mpc-ref">mpc 模块: <code>mpc-fl metrics</code></div>
<span class="arrow">→</span>
</a>
<a class="card" data-level="L2" href="/dp_primitives/">
<div class="head"><span class="num">16</span><span class="level">L2</span></div>
<h3><span class="ico">🛡️</span>差分隐私原语</h3>
<div class="name-en">mpc-privacy dp.{clip_l2, gaussian_noise, laplace_noise, GaussianAccountant}</div>
<div class="name-en">mpc-fl {clip_l2, gaussian_noise, GaussianAccountant}</div>
<div class="desc">裁剪 + 高斯/拉普拉斯噪声 + RDP 账本:脱敏直方图发布,输出 (ε, δ)-DP 保证。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-privacy dp</code></div>
<div class="mpc-ref">mpc 模块: <code>mpc-fl dp</code></div>
<span class="arrow">→</span>
</a>
</div>
@@ -347,9 +347,9 @@
<a class="card" data-level="L3" href="/aggregator_chain/">
<div class="head"><span class="num">17</span><span class="level">L3</span></div>
<h3><span class="ico">🔗</span>聚合器链</h3>
<div class="name-en">mpc-ml aggregators.{SecureSum, Clipping, Zeroing, Mean, GaussianNoise, build_chain}</div>
<div class="name-en">mpc-fl {SecureSum, Clipping, Zeroing, build_chain}</div>
<div class="desc">TFF 风格可插拔:Clipping → Zeroing → SecureSum → Mean → GaussianNoise 一键拼装。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-ml aggregators</code></div>
<div class="mpc-ref">mpc 模块: <code>mpc-fl aggregators</code></div>
<span class="arrow">→</span>
</a>
<a class="card" data-level="L3" href="/vframe_hframe/">
@@ -363,17 +363,17 @@
<a class="card" data-level="L3" href="/scql_explain/">
<div class="head"><span class="num">19</span><span class="level">L3</span></div>
<h3><span class="ico">📜</span>SCQL 查询计划</h3>
<div class="name-en">mpc-sql SecureSQL.explain</div>
<div class="name-en">mpc-sql Engine::explain</div>
<div class="desc">explain() 不执行只分析:哪些表/列/聚合/隐私约束会被使用;拒绝执行危险 SQL。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-sql SecureSQL.explain</code></div>
<div class="mpc-ref">mpc 模块: <code>mpc-sql Engine::explain</code></div>
<span class="arrow">→</span>
</a>
<a class="card" data-level="L3" href="/ole_paillier/">
<div class="head"><span class="num">20</span><span class="level">L3</span></div>
<h3><span class="ico">🔐</span>OLE / Paillier</h3>
<div class="name-en">mpc-crypto {paillier, ole}</div>
<div class="name-en">mpc-he {paillier, ole}</div>
<div class="desc">Paillier 加法同态(Enc(a)+Enc(b)=Enc(a+b))+ OLE 不经意线性求值。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-crypto {paillier, ole}</code></div>
<div class="mpc-ref">mpc 模块: <code>mpc-he {paillier, ole}</code></div>
<span class="arrow">→</span>
</a>
<a class="card" data-level="L2" href="/zkp/">
@@ -399,7 +399,7 @@
<h3><span class="ico">🛡️</span>安全合规检测</h3>
<div class="name-en">security_audit</div>
<div class="desc">站在可信数据空间流通运营方视角,对算法代码做静态合规检测:明文外泄 / 弱密码 / 密钥管理 / TEE 远程证明 / 隐私预算。每条命中点出风险等级、法规条款、修复建议与 mpc 替代方案。</div>
<div class="mpc-ref">规则库: <code>14 条静态 + 2 条 AST</code></div>
<div class="mpc-ref">规则库: <code>F01–F11 静态规则 + 14 条红线 + LLM 语义复核</code></div>
<span class="arrow">→</span>
</a>
<a class="card" data-level="L2" href="/secure_scenario/">
@@ -407,18 +407,13 @@
<h3><span class="ico">🔐</span>可信场景算法落地</h3>
<div class="name-en">secure_scenario</div>
<div class="desc">针对应用场景描述,依托 ssw-mpc + TEE 给出端到端可运行实现:威胁建模 → 算法流程 → 真实代码 → 执行结果 + 隐私预算 (ε,δ) + TEE 远程证明。4 个真实场景:流行病学 / 黑名单 / 联邦 / PIR。</div>
<div class="mpc-ref">mpc 模块: <code>mpc-psi / mpc-ml aggregators / mpc-crypto pir</code></div>
<div class="mpc-ref">mpc 模块: <code>mpc-psi / mpc-fl aggregators / mpc-he pir</code></div>
<span class="arrow">→</span>
</a>
</div>
</section>
<footer>
<div class="col">
<h4>本地启动</h4>
<div>启动: <code>python demo/server.py --host 0.0.0.0 --port 18850</code></div>
<div>入口: <code>http://localhost:18850/</code></div>
</div>
<div class="col">
<h4>算子级别</h4>
<div><span style="color:var(--l1)">● L1</span> 完整 7 阶段流程动画 + 数据表 + 算法对比</div>
@@ -439,17 +434,19 @@
</footer>
<script>
// 后端在线状态自检
// 后端在线状态自检 —— 读 /api/health 的真实值,不在前端编造版本号。
(async () => {
const dot = document.getElementById('status-dot');
try {
const r = await fetch('/psi/api/example');
const data = await r.json();
const h = await (await fetch('/api/health')).json();
document.getElementById('status-text').textContent = '后端在线';
document.getElementById('status-pid').textContent = data.meta ? '—' : '—';
document.getElementById('status-py').textContent = '3.13.12';
document.getElementById('status-backend-name').textContent = h.backend || '—';
document.getElementById('status-secure-val').textContent = h.secure ? 'on' : 'off';
// secure 关掉时标黄,避免「以为在跑安全模式」
if (!h.secure) document.getElementById('status-secure').style.color = 'var(--bob)';
} catch (e) {
document.getElementById('status-text').textContent = '后端离线';
document.getElementById('status-dot').style.background = 'var(--warn)';
dot.style.background = 'var(--warn)';
}
})();
</script>
+28 -2
View File
@@ -191,6 +191,32 @@
const $ = s => document.querySelector(s);
function esc(s){ return (s==null)?'':String(s).replace(/[&<>"']/g,c=>({'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;',"'":'&#39;'}[c])); }
/* 渲染「推理中」面板。
*
* 思考片段是长文本,外层提示是居中的 —— 继承居中会让一段英文推理被折成
* 居中几行,左右留大片空白,极难读。这里用 .thinking-panel 强制左对齐,
* 让它像日志一样从左往右顺读。
*
* `tail` 是后端保留的思考尾部(滚动窗口),所以渲染完自动滚到底部。
*/
function renderThinking(stage, tail) {
const t = (tail || '').replace(/[&<>]/g, c => ({'&':'&amp;','<':'&lt;','>':'&gt;'}[c]));
const n = t.length;
return `<div style="color:var(--text-dim);padding:24px 0;text-align:center;font-size:12px">
<span class="spinner"></span> ${esc(stage || '推理中')}…
<div class="thinking-panel">
<div class="tp-head"><span>模型思考过程</span><span class="tp-count">${n ? n + ' 字符' : ''}</span></div>
<div class="tp-body" id="tp-body">${t}</div>
</div>
</div>`;
}
/* 轮询刷新后把思考区滚到底部(tail 是尾部窗口,最新内容在最后)。 */
function scrollThinkingToBottom() {
const b = document.getElementById('tp-body');
if (b) b.scrollTop = b.scrollHeight;
}
let PRESET_DESC = {}, CURRENT = 'epidemiology', LAST_CODE = '';
const CHIPS = [
@@ -444,8 +470,8 @@ $('#btn-gen').onclick = async () => {
const job = await (await fetch(`/api/job/${j0.job_id}`)).json();
if (job.status === 'running') {
const p = job.progress || {};
const tail = (p.thinking_tail || '').replace(/[&<>]/g, c => ({'&':'&amp;','<':'&lt;','>':'&gt;'}[c]));
$('#understanding').innerHTML = `<div style="color:var(--text-dim);padding:24px 0;text-align:center;font-size:12px"><span class="spinner"></span> ${esc(p.stage || '推理中')}…<br/><span style="font-family:var(--mono);font-size:10px;color:var(--text-faint);display:block;margin-top:8px;max-height:150px;overflow:hidden;white-space:pre-wrap">${tail}</span></div>`;
$('#understanding').innerHTML = renderThinking(p.stage, p.thinking_tail);
scrollThinkingToBottom();
} else if (job.status === 'done') {
clearInterval(timer);
renderGen(job.result || {ok:false});
+29 -2
View File
@@ -186,6 +186,33 @@ const $ = s => document.querySelector(s);
const LEVELS = ['Critical','High','Medium','Low','Info'];
const LEVEL_COLOR = { Critical:'var(--warn)', High:'#fb923c', Medium:'var(--bob)', Low:'var(--alice)', Info:'var(--text-dim)' };
/* 渲染「推理中」面板。
*
* 思考片段是长文本,外层提示是居中的 —— 继承居中会让一段英文推理被折成
* 居中几行,左右留大片空白,极难读。这里用 .thinking-panel 强制左对齐,
* 让它像日志一样从左往右顺读。
*
* `tail` 是后端保留的思考尾部(滚动窗口),所以渲染完自动滚到底部,
* 免得每次轮询都把用户甩回中间。
*/
function renderThinking(stage, tail) {
const t = (tail || '').replace(/[&<>]/g, c => ({'&':'&amp;','<':'&lt;','>':'&gt;'}[c]));
const n = t.length;
return `<div style="color:var(--text-dim);padding:22px 0;text-align:center;font-size:12px">
<span class="spinner"></span> ${esc(stage || '推理中')}…
<div class="thinking-panel">
<div class="tp-head"><span>模型思考过程</span><span class="tp-count">${n ? n + ' 字符' : ''}</span></div>
<div class="tp-body" id="tp-body">${t}</div>
</div>
</div>`;
}
/* 轮询刷新后把思考区滚到底部(tail 是尾部窗口,最新内容在最后)。 */
function scrollThinkingToBottom() {
const b = document.getElementById('tp-body');
if (b) b.scrollTop = b.scrollHeight;
}
let TEMPLATES = {};
async function loadTpls() {
const r = await fetch('/security_audit/api/templates');
@@ -305,8 +332,8 @@ $('#btn-audit').onclick = async () => {
const job = await (await fetch(`/api/job/${j0.job_id}`)).json();
if (job.status === 'running') {
const p = job.progress || {};
const tail = (p.thinking_tail || '').replace(/[&<>]/g, c => ({'&':'&amp;','<':'&lt;','>':'&gt;'}[c]));
$('#result').innerHTML = `<div style="color:var(--text-dim);padding:22px 0;text-align:center;font-size:12px"><span class="spinner"></span> ${esc(p.stage || '推理中')}…<br/><span style="font-family:var(--mono);font-size:10px;color:var(--text-faint);display:block;margin-top:8px;max-height:140px;overflow:hidden;white-space:pre-wrap">${tail}</span></div>`;
$('#result').innerHTML = renderThinking(p.stage, p.thinking_tail);
scrollThinkingToBottom();
} else if (job.status === 'done') {
clearInterval(timer);
renderResult(job.result || {ok:false});
+43 -2
View File
@@ -117,12 +117,21 @@ fn digit_runs(line: &str) -> (usize, bool) {
}
/// 硬编码密钥:`api_key = "……"` 且值不短于 12 位字母数字。
///
/// # 必须兼容 Rust 的类型标注
///
/// Python 写 `API_KEY = "sk-live-..."`,Rust 写 `const API_KEY: &str = "sk-live-..."`
/// —— 等号前夹着 `: &str`。只按「名字后面紧跟 `=`」匹配,Rust 这种**最常见**的
/// 硬编码写法会整个漏掉(这正是本规则的真实盲区)。
/// 所以这里允许在名字与等号之间跳过一段 Rust 类型标注。
fn hardcoded_secret(line: &str) -> bool {
let l = line.to_ascii_lowercase();
for name in ["api_key", "api-key", "apikey", "secret", "token", "passwd", "password"] {
let Some(pos) = l.find(name) else { continue };
let rest = l[pos + name.len()..].trim_start();
let Some(rest) = rest.strip_prefix('=') else { continue };
// `skip_type_annotation` 返回**等号之后**的内容。
let Some(rest) = skip_type_annotation(&l[pos + name.len()..]) else {
continue;
};
let rest = rest.trim_start();
let Some(quote) = rest.chars().next().filter(|c| *c == '"' || *c == '\'') else {
continue;
@@ -139,6 +148,38 @@ fn hardcoded_secret(line: &str) -> bool {
false
}
/// 跳过「名字」与「=` 之间可能的 Rust 类型标注,返回**等号之后**的切片。
///
/// 认得的形式:
/// - `API_KEY = "..."`(Python,无标注)→ 直接取 `=` 之后
/// - `const API_KEY: &str = "..."` → 跳过 `: &str`
/// - `let api_key: String = "..."` → 跳过 `: String`
///
/// 标注只允许类型里会出现的字符,且长度有上限 —— 否则容易跳到别的语句里去。
fn skip_type_annotation(rest: &str) -> Option<&str> {
let trimmed = rest.trim_start();
// 无标注:直接是 `=`
if let Some(r) = trimmed.strip_prefix('=') {
return Some(r);
}
// 有标注:必须是 `:` 开头,且整段只含类型字符
let after_colon = trimmed.strip_prefix(':')?;
for (i, c) in after_colon.char_indices() {
if c == '=' {
return Some(&after_colon[i + 1..]);
}
let ok = c.is_alphanumeric()
|| matches!(c, '_' | '&' | '*' | '<' | '>' | '[' | ']' | ',' | '(' | ')' | ' ' | ':');
if !ok {
return None;
}
if i > 48 {
return None;
}
}
None
}
/// `let _ = ...` 是否真的**丢掉了返回值**(吞异常的典型写法)。
///
/// 只丢掉一个 `Result` 才算吞异常;`let _ = (sha256, derive, Stream::new);`
+112 -37
View File
@@ -698,47 +698,76 @@ fn job_poll(ctx: &Ctx) -> Result<Json, String> {
// ─────────────────────────── 案例模板 ───────────────────────────
const BAD_CODE: &str = r#"import hashlib, json, pickle, random, subprocess
import pandas as pd, requests
/// 反面案例:**Rust** 写的错误实现(ssw-mpc / Rust 风格)。
///
/// 为什么必须是 Rust 而不是 Python:本平台是 Rust 框架,审计引擎收到的
/// 待审代码绝大多数是 Rust。拿一份 Python 反例当默认样例,会让人误以为
/// 「平台审的是 Python」,也让规则引擎的 Rust 分支(方法调用、Rust 专有
/// 弱哈希、`let _ =` 吞异常)失去演示意义。
///
/// 这份代码刻意覆盖 F01–F11 全部规则,且**故意用错 API 签名**
/// (`blind` 传 4 参、`match_across` 传 4 参、`Handle::new` 少传 Label)
/// —— 真实模型生成里最常见、也最难自查出来的错误。
const BAD_CODE: &str = r#"//! 两家医院联合统计糖尿病阳性率 · 错误实现(反面案例)。
//!
//! 典型毛病:明文 join / 明文落盘 / md5 假匿名 / 假 DP / eval 注入 /
//! 无远程证明 / 吞异常 / API 签名写错。
# 从两家医院拉全量患者明细做联合统计
patients = pd.read_csv("/data/hospital_a_patients.csv") # 含身份证、手机号、诊断
peer = pd.read_csv("/data/hospital_b_patients.csv")
use mpc_core::hash::sha256;
use mpc_core::prg::Stream;
API_KEY = "sk-live-9f3c1a77b2e04d6890ffab12cd34ef56" # 直连第三方分析服务
use mpc_fl::{clip_l2, gaussian_noise};
use mpc_policy::{Guard, Handle, Label};
use mpc_psi::{blind, match_across, reblind};
# 用身份证号做主键碰撞,直接明文 join
merged = patients.merge(peer, on="id_card", how="inner")
print("联合后的患者明细:", merged.head(20).to_dict())
// 硬编码第三方服务密钥(应从 KMS 注入)
const API_KEY: &str = "sk-live-9f3c1a77b2e04d6890ffab12cd34ef56";
# 口令与匿名 id 都用 md5
merged["pseudo_id"] = merged["id_card"].apply(lambda s: hashlib.md5(s.encode()).hexdigest())
admin_pwd = hashlib.md5(b"Hosp!2026").hexdigest()
fn main() {
// 从两家医院拉全量患者明细做联合统计(含身份证、手机号、诊断)
let patients = std::fs::read_to_string("/data/hospital_a_patients.csv").unwrap();
let peer = std::fs::read_to_string("/data/hospital_b_patients.csv").unwrap();
# 结果落盘 + 回传第三方
merged.to_csv("/tmp/joint_patients.csv", index=False)
with open("/tmp/cache.pkl", "wb") as f:
pickle.dump(merged, f)
requests.post("https://analytics.example.com/upload",
json={"key": API_KEY, "rows": merged.to_dict("records")})
// 用身份证号做主键碰撞,直接明文 join
let merged = format!("{patients}\n{peer}");
println!("联合后的患者明细: {}", &merged[..120]);
# 号称"差分隐私",实际只是随便加了个噪声,没有任何预算记账
positive = int(merged["diabetes"].sum()) + random.gauss(0, 1.0)
// 口令与匿名 id 都用 md5
let pseudo_id = md5::compute(b"310101199001011234");
let admin_pwd = md5::compute(b"Hosp!2026");
# 运营方传进来的过滤条件直接拼进查询并 eval
flt = get_user_filter()
subset = eval(f"merged[merged['age'] {flt}]", {"merged": merged})
subprocess.run(f"rsync -a /tmp/joint_patients.csv backup:{flt}", shell=True)
// 结果落盘 + 回传第三方
std::fs::write("/tmp/joint_patients.csv", merged.as_bytes()).unwrap();
let _cache = std::fs::write("/tmp/cache.pkl", merged.as_bytes()).unwrap();
let _client = reqwest::Client::new();
let _resp = reqwest::Client::new().post("https://analytics.example.com/upload").body(merged.clone()).send();
# 模型侧:不做远程证明就把明文推进 enclave
enclave.load_plaintext(merged.values)
// 号称"差分隐私",实际只是随便加了个噪声,没有任何预算记账
let positive = 42.0 + gaussian_noise(1, 1.0, 7)[0];
try:
verify_partner_signature(peer)
except:
pass # 签名验不过也继续跑
// 运营方传进来的过滤条件直接拼进查询并 eval
let flt = get_user_filter();
let subset = eval(f"merged[merged['age'] {flt}]");
let _cmd = std::process::Command::new("/bin/sh").arg("-c").arg(format!("rsync -a /tmp/joint_patients.csv backup:{flt}")).spawn();
print(f"阳性人数(加噪) = {positive}, 身份证样例 = 310101199001011234")
// 模型侧:不做远程证明就把明文推进 enclave
let _plain = enclave.load_plaintext(merged.as_bytes());
// 签名验不过也继续跑
if let Err(_e) = verify_partner_signature(&peer) {
let _ = std::fs::write("/tmp/verify_failed.log", b"ignored");
}
// API 签名写错:blind 传了 4 个参数(真实签名 3 个),
// match_across 传了 4 个(真实 2 个),Handle::new 少传 Label
let key_a = sha256(b"psi-key-a");
let mut stream = Stream::new(&key_a);
let _pts = blind(&[], &key_a, &mut stream, "extra-domain");
let _hits = match_across(&[], &[], &key_a, &key_b);
let _handle = Handle::new("hospital_a");
println!("阳性人数(加噪) = {positive}, 身份证样例 = 310101199001011234");
}
"#;
const GOOD_CODE: &str = r#"//! 两家医院联合统计糖尿病阳性率 · 可信数据空间合规实现(ssw-mpc / Rust)。
@@ -852,7 +881,7 @@ fn templates(_ctx: &Ctx) -> Result<Json, String> {
crate::jobj! {
"id" => "bad_v1",
"name" => "反面案例 · 联合统计的错误实现",
"desc" => "明文 join / 明文回传 / md5 假匿名 / 假 DP / eval 注入 / 无远程证明",
"desc" => "Rust 写法 · 明文 join / 明文落盘回传 / md5 假匿名 / 假 DP / eval 注入 / 无远程证明 / 吞异常 / API 签名写错",
"code" => BAD_CODE,
},
crate::jobj! {
@@ -886,18 +915,26 @@ mod tests {
);
}
/// 反面案例(Python)应命中全套 F01–F10。
/// 反面案例(Rust)应命中全套 F01–F11。
#[test]
fn bad_python_example_hits_all_rules() {
fn bad_rust_example_hits_all_rules() {
let found = ids(BAD_CODE);
for id in ["F01", "F02", "F03", "F04", "F05", "F06", "F07", "F09", "F10"] {
for id in ["F01", "F02", "F03", "F04", "F05", "F06", "F07", "F08", "F09", "F10", "F11"] {
assert!(found.contains(&id), "反面案例漏判 {id},实际命中 {found:?}");
}
}
/// Rust 反例:明文落盘 + 弱哈希 + 外发未证明 + 加噪未记账 + 吞异常。
/// 反面案例里故意写错的 API 签名必须被 F11 抓到 —— 这是「代码编译不过」的
/// 唯一可机检信号,也是平台自生成代码最常见的翻车点。
#[test]
fn bad_rust_example_hits_rules() {
fn bad_rust_example_catches_wrong_signatures() {
let found = ids(BAD_CODE);
assert!(found.contains(&"F11"), "反面案例里的错签名没被 F11 抓到");
}
/// 最小 Rust 反例:明文落盘 + 弱哈希 + 外发未证明 + 加噪未记账 + 吞异常。
#[test]
fn minimal_bad_rust_snippet_hits_rules() {
let code = r#"
fn main() {
let rows = std::fs::write("/tmp/plain.csv", b"id_card,name").unwrap();
@@ -1098,6 +1135,44 @@ fn run(mut guard: Guard, tee: &mut dyn Tee) -> Result<(), String> {
);
}
/// F04 必须认 Rust 的**类型标注写法**:`const API_KEY: &str = "..."`。
///
/// 这是本规则的真实盲区:原来只按「名字后面紧跟 `=`」匹配,Rust 最常见的
/// 硬编码密钥写法(中间夹着 `: &str`)会整个漏掉 —— 等号前有标注就匹配不上。
#[test]
fn f04_detects_rust_const_with_type_annotation() {
let rust_const = r#"
const API_KEY: &str = "sk-live-9f3c1a77b2e04d6890ffab12cd34ef56";
fn main() {}
"#;
assert!(
ids(rust_const).contains(&"F04"),
"Rust const + 类型标注的硬编码密钥漏判(盲区未修)"
);
let rust_let = r#"
fn main() {
let api_key: String = "sk-live-9f3c1a77b2e04d6890ffab12cd34ef56".to_string();
}
"#;
assert!(
ids(rust_let).contains(&"F04"),
"Rust let + 类型标注的硬编码密钥漏判(盲区未修)"
);
// Python 写法仍要认
assert!(
ids(r#"API_KEY = "sk-live-9f3c1a77b2e04d6890ffab12cd34ef56""#).contains(&"F04"),
"Python 硬编码密钥漏判"
);
// 短值 / 非密钥名不该误报
assert!(
!ids(r#"const NAME: &str = "hospital";"#).contains(&"F04"),
"普通常量被误判成硬编码密钥"
);
}
/// F09 认 Rust 的 `AuditLog` / `audit.record`。
#[test]
fn f09_detects_rust_audit() {