44-EFT.WP.Data.ModelCards v1.0 | 第14章 公平性、伦理与使用限制

目录 ／ 文档-技术白皮书 ／ 44-EFT.WP.Data.ModelCards v1.0

第14章 公平性、伦理与使用限制

I. 章节目的与范围

II. 术语与依赖

• 依赖卷：数据契约/导出《EFT.WP.Core.DataSpec v1.0》；计量与量纲校核《EFT.WP.Core.Metrology v1.0》；数据事实与覆盖/切分《EFT.WP.Data.DatasetCards v1.0》（隐私、伦理与合规）。
• 数学与符号：公平性差异度量采用 gap_metric（如 abs_diff 或 ratio）；内联符号用反引号（如 Δ_gap），含除号/积分/复合算符必须加括号；公式/符号/定义禁用中文。

III. 字段与结构（规范性）

fairness:

axes: ["class","region","device","language"] # 评估轴

gap_metric: "abs_diff|ratio" # 差异度量

threshold: 0.05 # 公平性阈值

stratification: ["val","test"] # 分层集合

mitigation:

enabled: true

methods: ["reweight","resample","calibration","post-hoc-threshold"]

reeval_required: true

reporting:

include_ci: true # 指标配 95% CI

table_axes: ["axis","bucket","metric"]

significance: {test:"bootstrap", alpha:0.05}

ethics:

intended_use: ["academic","benchmark","safety-research"] # 允许用途

prohibited_use: ["surveillance","biometric_identification","unlawful_discrimination"]

disclosures:

sensitive_attributes: ["N/A"] # 如涉及则列出处理方式

human_in_the_loop: true

risk_notes: "Model outputs must be reviewed in high-stakes contexts."

governance:

review_process: ["internal-ethics-board"]

update_policy: "on-drift|on-incident|quarterly"

usage:

regional_compliance: ["EU-GDPR","US-CCPA"] # 区域/法规限制

access_control:

roles: ["owner","maintainer","reader"]

enforcement: ["signed-url","token","ip-allowlist"]

rate_limits:

qps_max: 1000

burst: 200

monitoring:

online_checks: ["fairness_gap","drift_kl","error_rate"]

alert_rules:

- {name:"fairness_gap_breach", rule:"Δ_gap>0.05 for 60m", severity:"high"}

IV. 公平性评测与阈值

1. 轴与分桶：axes 明确评估维度（类别/区域/设备/语言等），与数据卡 coverage 的分层映射一致；每个桶需报告样本数与 95% 置信区间。
2. 差异度量：
   • 绝对差：Δ_gap = ( metric_ref - metric_grp )；
   • 比值差：Δ_gap = ( metric_grp / metric_ref )；
     超阈 threshold 视为阻断项或需给出缓解与再评测结果。
3. 缓解与复评：记录 mitigation.methods（重加权/重采样/校准/后处理阈值）与 reeval_required=true；复评遵循第11章协议（seeds/repeats/ci/significance）。

V. 伦理披露与人机协同

• 用途声明：intended_use/prohibited_use 明确边界；高风险场景须启用 human_in_the_loop=true 并记录人工复核环节。
• 敏感属性：如涉及敏感属性，应声明采集/处理/脱敏方式与限制；禁止对未获同意的敏感属性进行推断或用于决策。
• 治理流程：governance.review_process 记录内部伦理评审；更新策略 update_policy 与事件触发条件明确。

VI. 使用限制与线上监测

• 区域合规：regional_compliance 与数据卡一致；涉及跨境传输需在导出清单提供相应机制/模板编号。
• 访问与速率：access_control 与 rate_limits 明确；生产监测使用 monitoring.online_checks 与 alert_rules，超阈需触发降级或阻断。

VII. 计量与单位（如涉性能/时延/能耗）

VIII. 机器可读片段（可直接嵌入）

fairness:

axes: ["class","region","device"]

gap_metric: "abs_diff"

threshold: 0.05

stratification: ["test"]

mitigation: {enabled:true, methods:["reweight","calibration"], reeval_required:true}

reporting: {include_ci:true, table_axes:["axis","bucket","metric"], significance:{test:"bootstrap", alpha:0.05}}

ethics:

intended_use: ["academic","benchmark"]

prohibited_use: ["surveillance","biometric_identification"]

disclosures: {sensitive_attributes:["N/A"], human_in_the_loop:true, risk_notes:"Human review required in high-stakes."}

governance: {review_process:["internal-ethics-board"], update_policy:"on-drift"}

usage:

regional_compliance: ["EU-GDPR"]

access_control: {roles:["owner","maintainer","reader"], enforcement:["signed-url","token"]}

rate_limits: {qps_max: 1000, burst: 200}

monitoring:

online_checks: ["fairness_gap","drift_kl"]

alert_rules: [{name:"fairness_gap_breach", rule:"Δ_gap>0.05 for 60m", severity:"high"}]

IX. 与评测协议、训练数据与部署的一致性

• 公平性评测使用冻结切分与分层映射（映射到数据卡 coverage）；
• 若采用缓解策略（重加权/后处理阈值），需在 optimization/hyperparams 与 evaluation 同步记录并做显著性检验；
• 线上监测与部署端点定义在第16章 API 绑定中给出 OpenAPI 片段。

X. 导出清单与审计轨

export_manifest:

artifacts:

- {path:"fairness/by_axis_metrics.csv", sha256:"..."}

- {path:"fairness/mitigation_report.md", sha256:"..."}

- {path:"usage/alert_rules.yaml", sha256:"..."}

references:

- "EFT.WP.Core.DataSpec v1.0:EXPORT"

- "EFT.WP.Core.Metrology v1.0:check_dim"

- "EFT.WP.Data.DatasetCards v1.0:Ch.13"

XI. 本章合规自检

• axes/gap_metric/threshold 明确且与数据卡分层一致；各桶报告样本数与 95% CI，并进行显著性检验。
• 触发阈值后已给出缓解方案并复评；结果达标或标注剩余风险与使用限制。
• 伦理披露与人机协同流程已记录；禁止用途清晰且在部署层强制。
• 区域合规、访问控制与线上监测规则完备；性能/能耗等单位通过 check_dim。
• 导出清单列出表格/报告/告警配置并具 sha256；引用携带“卷名 vX.Y:锚点”。