{
  "format": "agent-article-v0",
  "article_id": "EA-080",
  "canonical_url": "/tools/ea-080-synthetic-data",
  "page_type": "reference",
  "claim_ids": [],
  "evidence_ids": [
    "EVD-EA-080-URL-95F8EBF0F0DF",
    "EVD-EA-080-URL-D316DBE7B2D1",
    "EVD-EA-080-URL-86E4468D1454",
    "EVD-EA-080-URL-B778EC39D8DB",
    "EVD-EA-080-URL-B6DAB093A0B5",
    "EVD-EA-080-URL-B6D243A55F78",
    "EA-080-R11-RESEARCH",
    "EA-080-R11-WEB-READBACK"
  ],
  "unknown_ids": [
    "EA-080-UNKNOWN-01",
    "EA-080-UNKNOWN-CURRENTNESS"
  ],
  "operational_object_ids": [],
  "provider": [],
  "models": [],
  "plans": [],
  "surfaces": [],
  "regions": [],
  "current_status": "UNKNOWN",
  "observation_status": "SPEC_ONLY",
  "freshness": "UNKNOWN",
  "last_verified": null,
  "supersedes": [],
  "superseded_by": [],
  "provenance": [
    "COMM-075-EA0080",
    "SRC-EA-080-URL-95F8EBF0F0DF",
    "SRC-EA-080-URL-D316DBE7B2D1",
    "SRC-EA-080-URL-86E4468D1454",
    "SRC-EA-080-URL-B778EC39D8DB",
    "SRC-EA-080-URL-B6DAB093A0B5",
    "SRC-EA-080-URL-B6D243A55F78",
    "WEBBODY-EA0080-R8-DB363647"
  ],
  "article_sources": [
    {
      "id": "EVD-EA-080-URL-95F8EBF0F0DF",
      "evidence_class": "RESEARCH",
      "title": "| official guidance | Guidelines for Evaluating Differential Privacy Guarantees / NIST |",
      "href": "https://www.nist.gov/publications/guidelines-evaluating-differential-privacy-guarantees",
      "supports": [],
      "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。",
      "origin_class": "SOURCE_PACKET_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-EA-080-URL-95F8EBF0F0DF"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "| official guidance | Guidelines for Evaluating Differential Privacy Guarantees / NIST | https://www.nist.gov/publications/guidelines-evaluating-differential-privacy-guarantees"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-EA-080-URL-D316DBE7B2D1",
      "evidence_class": "RESEARCH",
      "title": "| official project | Differential Privacy / NIST Privacy Engineering Collaboration Space |",
      "href": "https://csrc.nist.gov/projects/privacy-engineering/collaboration-space/focus-areas/de-id/tools",
      "supports": [],
      "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。",
      "origin_class": "SOURCE_PACKET_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-EA-080-URL-D316DBE7B2D1"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "| official project | Differential Privacy / NIST Privacy Engineering Collaboration Space | https://csrc.nist.gov/projects/privacy-engineering/collaboration-space/focus-areas/de-id/tools"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-EA-080-URL-86E4468D1454",
      "evidence_class": "RESEARCH",
      "title": "| author-maintained OSS | Synthetic Data Vault (SDV) / DataCebo |",
      "href": "https://github.com/sdv-dev/SDV",
      "supports": [],
      "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。",
      "origin_class": "SOURCE_PACKET_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-EA-080-URL-86E4468D1454"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "| author-maintained OSS | Synthetic Data Vault (SDV) / DataCebo | https://github.com/sdv-dev/SDV"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-EA-080-URL-B778EC39D8DB",
      "evidence_class": "RESEARCH",
      "title": "| primary/benchmark research | Anonymeter: Anonymity and Privacy Risk Metrics for Relational Data / Giomi et al. |",
      "href": "https://arxiv.org/abs/2211.10459",
      "supports": [],
      "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。",
      "origin_class": "SOURCE_PACKET_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-EA-080-URL-B778EC39D8DB"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "| primary/benchmark research | Anonymeter: Anonymity and Privacy Risk Metrics for Relational Data / Giomi et al. | https://arxiv.org/abs/2211.10459"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-EA-080-URL-B6DAB093A0B5",
      "evidence_class": "RESEARCH",
      "title": "| research | Really Useful Synthetic Data — A Framework To Evaluate The Quality Of Differentially Private Synthetic Data / Garrido et al. |",
      "href": "https://arxiv.org/abs/2306.05657",
      "supports": [],
      "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。",
      "origin_class": "SOURCE_PACKET_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-EA-080-URL-B6DAB093A0B5"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "| research | Really Useful Synthetic Data — A Framework To Evaluate The Quality Of Differentially Private Synthetic Data / Garrido et al. | https://arxiv.org/abs/2306.05657"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EVD-EA-080-URL-B6D243A55F78",
      "evidence_class": "RESEARCH",
      "title": "| research | Membership Inference Attacks Against Language Models Via Neighbourhood Comparison / Mattern et al. |",
      "href": "https://arxiv.org/abs/2305.18462",
      "supports": [],
      "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。",
      "origin_class": "SOURCE_PACKET_REFERENCE",
      "source_or_experiment_ids": [
        "SRC-EA-080-URL-B6D243A55F78"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "| research | Membership Inference Attacks Against Language Models Via Neighbourhood Comparison / Mattern et al. | https://arxiv.org/abs/2305.18462"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EA-080-R11-RESEARCH",
      "evidence_class": "RESEARCH",
      "title": "EA-080 監査済みresearch / canonical revision binding",
      "supports": [
        "1. Synthetic dataは「元データを別の値へ置換したもの」という一種類ではない。statistical/tabular generator、simulation、rule-based generator、foundation-model/LLM generation、DP mechanismを伴う生成など、生成経路と保証が異なる。したがってartifactへ generator class/version/config とsource cohortを残す必要がある。\n2. DPはprivacy budgetとmechanismを明示した場合に形式的なprivacy guaranteeを論じられるが、「synthetic」というラベルだけではDP保証にならない。DP使用時もepsilon/delta、adjacency notion、composition、clipping/bounds、randomness、accounting、implementationを記録対象とする。\n3. privacyとutilityは別軸である。distributional similarityが高いほど安全とは限らず、privacy riskが低くてもdownstream taskで使えるとは限らない。単一synthetic-quality scoreでrelease判断しない。\n4. downstream validityも独立軸である。synthetic上で得たモデル順位・fairness・rare-event performance・calibrationがreal holdoutへ移るかは、real holdoutを使った外部検証なしには確定できない。\n5. LLM/text generationでは、verbatim/near-verbatim reproduction、PII/secret reconstruction、membership/attribute inferenceを別failure surfaceとする。promptで「匿名化して」と命令しただけではprivacy mechanismの証明にならない。\n6. rare group/rare eventは二重の問題を持つ。過度に平滑化すればutility/fairnessを失い、忠実に再現しすぎればunique/rare recordのprivacy riskを高め得る。minority preservationとprivacy riskを同時監査する。"
      ],
      "limit": "canonical監査・revision bindingの存在を示す。外部ソースcurrentnessが本文またはcontrol-planeで確定していない場合は、currentnessをUNKNOWNのまま保持する。",
      "origin_class": "AUDITED_CANONICAL_RESEARCH_PACKET",
      "source_or_experiment_ids": [
        "COMM-075-EA0080"
      ],
      "supporting_locators": [
        "COMM-075-EA0080",
        "EA-080-R5-LEGACY-MIGRATION-20261008-R1",
        "content/staging/EA-080.md"
      ],
      "scope": "監査済みEA本文・canonical revision・既存Source Graph relationの範囲。欠損currentnessは推測で閉じない。",
      "provenance_status": "AUDIT_PASS_CANONICAL_BINDING"
    },
    {
      "id": "EA-080-R11-WEB-READBACK",
      "evidence_class": "IMPL",
      "title": "EA-080 Web body identity / canonical readback",
      "supports": [
        "canonical staging bodyのSHA-256一致とR9 canonical ledger readbackをProduction projectionの入力として検証する。"
      ],
      "limit": "body identity/readbackは本文claim自体の真偽や外部ソースの最新性を追加保証しない。",
      "checked_at": null,
      "origin_class": "IMPLEMENTATION_READBACK",
      "source_or_experiment_ids": [
        "WEBBODY-EA0080-R8-DB363647"
      ],
      "supporting_locators": [
        "WEBBODY-EA0080-R8-DB363647",
        "content/staging/EA-080.md",
        "EA-080-R5-LEGACY-MIGRATION-20261008-R1"
      ],
      "scope": "R9/R10 canonical readback後のProduction生成時body identity。",
      "provenance_status": "CANONICAL_READBACK_BODY_HASH_MATCH"
    }
  ],
  "research_sections": [
    {
      "heading": "WHY_THIS_MATTERS_TO_UPDATE",
      "text": "生成AI・分析AIの開発では、実データを直接複製せずに開発・評価・共有用データを作る要求が増える。一方、synthetic dataは「人工生成だから匿名」とは限らず、utility・fidelity・privacy leakage・rare-case preservation・downstream validityを別々に検証する必要がある。Enterprise AI Coreでは、生成器そのものよりも、source cohort→generator→privacy mechanism→release artifact→downstream evaluationのprovenanceとrelease gateを定義する価値が高い。"
    },
    {
      "heading": "SCOPE_ROUTING",
      "text": "Enterprise AI Core / governance / privacy / benchmark-evaluation / implementation / interoperability / non-LLM component。Provider固有のsynthetic-data productはprovider mappingへ分離する。"
    },
    {
      "heading": "CLAIM_OR_RESEARCH_QUESTION",
      "text": "1. Synthetic dataは実データの単純な匿名化・マスキングと同義ではないか。\n2. Differential Privacy (DP) を用いる場合、どの保証を形式化でき、何を保証できないか。\n3. Utility/fidelityとprivacyを同一スコアへ潰さず、release前にどの軸を独立評価すべきか。\n4. LLMでtabular/text synthetic dataを作る場合も、memorization・membership/attribute inference・rare-record leakageを独立failure surfaceとして扱うべきか。"
    },
    {
      "heading": "EVIDENCE TABLE",
      "text": "| source class | exact title / author-org | canonical URL | publication/update/access | version/release | supporting locator / claim | evidence label | relation |\n|---|---|---|---|---|---|---|---|\n| official guidance | Guidelines for Evaluating Differential Privacy Guarantees / NIST | https://www.nist.gov/publications/guidelines-evaluating-differential-privacy-guarantees | published 2025-03-19; accessed 2026-09-16 | NIST SP 800-226 | DPを評価するためのprivacy guarantee・threat model・parameter/implementation評価の枠組み | VERIFIED_SPEC / SOURCE_REPORT | support/context |\n| official project | Differential Privacy / NIST Privacy Engineering Collaboration Space | https://csrc.nist.gov/projects/privacy-engineering/collaboration-space/focus-areas/de-id/tools | accessed 2026-09-16 | current page | de-identification/privacy engineering tool landscape。単なる識別子除去とformal privacy guaranteeを混同しない背景 | SOURCE_REPORT | context/limit |\n| author-maintained OSS | Synthetic Data Vault (SDV) / DataCebo | https://github.com/sdv-dev/SDV | accessed 2026-09-16 | current upstream; exact release to be re-pinned at update work | tabular synthetic data generation、metadata/modeling/evaluation ecosystemが実装として存在 | OSS_PRACTICE / PRODUCT_VOLATILE | support |\n| primary/benchmark research | Anonymeter: Anonymity and Privacy Risk Metrics for Relational Data / Giomi et al. | https://arxiv.org/abs/2211.10459 | 2022; accessed 2026-09-16 | paper | singling-out/linkability/inference等の攻撃ベースprivacy risk評価を提案 | RESEARCH | support/limit |\n| research | Really Useful Synthetic Data — A Framework To Evaluate The Quality Of Differentially Private Synthetic Data / Garrido et al. | https://arxiv.org/abs/2306.05657 | 2023; accessed 2026-09-16 | paper | DP synthetic dataをutility/fidelity/privacy等の複数軸で評価する必要性 | RESEARCH | support |\n| research | Membership Inference Attacks Against Language Models Via Neighbourhood Comparison / Mattern et al. | https://arxiv.org/abs/2305.18462 | 2023; accessed 2026-09-16 | paper | language-model output/model accessに対するmembership inference riskの実証経路。LLM生成だからsource membership riskが自動消滅するわけではない | RESEARCH | counter/limit |"
    },
    {
      "heading": "FINDINGS",
      "text": "1. Synthetic dataは「元データを別の値へ置換したもの」という一種類ではない。statistical/tabular generator、simulation、rule-based generator、foundation-model/LLM generation、DP mechanismを伴う生成など、生成経路と保証が異なる。したがってartifactへ generator class/version/config とsource cohortを残す必要がある。\n2. DPはprivacy budgetとmechanismを明示した場合に形式的なprivacy guaranteeを論じられるが、「synthetic」というラベルだけではDP保証にならない。DP使用時もepsilon/delta、adjacency notion、composition、clipping/bounds、randomness、accounting、implementationを記録対象とする。\n3. privacyとutilityは別軸である。distributional similarityが高いほど安全とは限らず、privacy riskが低くてもdownstream taskで使えるとは限らない。単一synthetic-quality scoreでrelease判断しない。\n4. downstream validityも独立軸である。synthetic上で得たモデル順位・fairness・rare-event performance・calibrationがreal holdoutへ移るかは、real holdoutを使った外部検証なしには確定できない。\n5. LLM/text generationでは、verbatim/near-verbatim reproduction、PII/secret reconstruction、membership/attribute inferenceを別failure surfaceとする。promptで「匿名化して」と命令しただけではprivacy mechanismの証明にならない。\n6. rare group/rare eventは二重の問題を持つ。過度に平滑化すればutility/fairnessを失い、忠実に再現しすぎればunique/rare recordのprivacy riskを高め得る。minority preservationとprivacy riskを同時監査する。"
    },
    {
      "heading": "LIMITS / COUNTEREVIDENCE / FAILURE ENVELOPE",
      "text": "- Synthetic ≠ anonymous ≠ DP。\n- DP ≠ zero privacy risk。保証は定義したneighboring relation・mechanism・budget・implementationに依存する。\n- High statistical fidelity ≠ safe release。\n- Low attack success on one attack suite ≠ universal privacy proof。attacker knowledge/threat modelが変われば結果も変わる。\n- Synthetic-trained modelの良好なsynthetic validation ≠ real-world performance。real holdoutでの外部妥当性が必要。\n- Aggregate distribution一致はrare subgroup、tail、causal relation、longitudinal dependency、referential integrityの保存を保証しない。\n- OSS benchmark/metricの結果を全dataset・全generatorへ一般化しない。"
    },
    {
      "heading": "AVAILABILITY_GATES",
      "text": "- source data classification / legal basis / purpose limitation\n- generator type / version / configuration / seed\n- DP有無。DPならepsilon/delta、accountant、composition、bounds/clipping、neighbor definition\n- structured/tabular/text/image/time-series等のmodality\n- relational integrity / longitudinal identity requirement\n- tenant/region/data residency、training retention、logging\n- release audience（internal dev / vendor / public）とattacker capability\n- real holdoutを評価専用に保持できるか"
    },
    {
      "heading": "IMPLEMENTATION_STATE",
      "text": "実装可能。Enterprise実装は generator と release gate を分離する。推奨pipelineは `classified source → bounded cohort → generator/versioned config → synthetic artifact → utility/fidelity eval → privacy attack eval → policy release gate → downstream real-holdout validation`。LLMはgenerator候補の一つに過ぎず、privacy evaluatorやrelease authorityを兼任させない。"
    },
    {
      "heading": "BENCHMARK_OR_EVAL",
      "text": "後続TEST_PROFILE:\n- fixtures: balanced tabular / rare subgroup / high-cardinality categorical / correlated quasi-identifiers / longitudinal / relational / free-text with seeded canary PII\n- utility: univariate/multivariate distance、correlation/association、constraint validity、referential integrity、TSTR/TRTS（Train Synthetic Test Real / Train Real Test Synthetic）、real-holdout task metric、calibration\n- privacy: exact/near duplicate rate、nearest-neighbor distance、singling-out、linkability、inference、membership inference、canary extraction、PII/secret scanner\n- fairness: subgroup utility、error parity、rare-group preservation\n- DP: epsilon/delta/accounting validation、repeat-run budget composition\n- operability: generation latency、cost/record、reproducibility、audit completeness\n評価結果はprivacy/utility/fairness/operabilityを別列で保持し、加重平均一個へ圧縮しない。"
    },
    {
      "heading": "SECURITY/PRIVACY",
      "text": "raw sourceへのアクセス制御とsynthetic artifactのrelease classificationを別管理する。synthetic artifactを自動的にpublic扱いしない。canary/secret seeded testは実秘密情報を使わずsynthetic fixtureで行う。\nAUDIT: source dataset ID/hash、cohort query、schema、generator/version/hash、config/seed、DP parameters、evaluation suite/version、attack assumptions、release decision、approverを保存。"
    },
    {
      "heading": "COST",
      "text": "generator computeだけでなく、privacy attack suite、real-holdout evaluation、人手review、再生成を含める。"
    },
    {
      "heading": "OPERABILITY",
      "text": "source schema drift、generator upgrade、DP budget change、new attack/evaluator release時に再評価する。"
    },
    {
      "heading": "SAFE_WORDING_FOR_FUTURE_BOOK",
      "text": "「Synthetic dataは実データへの直接アクセスを減らし、開発・試験・共有の選択肢を増やせる。ただし“synthetic”という生成由来だけでは匿名性やprivacyを保証しない。Enterprise運用では、生成器・source cohort・privacy mechanismをversion固定し、utility、downstream validity、privacy attack、fairnessを独立評価してrelease gateを通す。Differential Privacyを使う場合は、epsilon/delta等の保証条件とaccountingをartifact provenanceとして保持する。」"
    },
    {
      "heading": "UPDATE_WORK_HANDOFF",
      "text": "- Core候補名: `SYNTHETIC DATA RELEASE CONTRACT`\n- 必須field候補: source dataset ID/hash; source classification; cohort query; schema/version; modality; generator class/product/model/version/hash; config/seed; training cutoff; DP mechanism; epsilon/delta; accountant/composition; constraints; artifact hash; fidelity metrics; downstream real-holdout metrics; privacy attack suite/version/results; subgroup metrics; PII/canary scan; release audience; retention; approver; provenance.\n- provider mapping時はcommercial synthetic-data productsのcurrent plan/region/retention/training-use/APIを再確認する。\n- EA-076 Time-series/Sensor AIとはlongitudinal synthetic telemetryで接続するが、時系列予測品質とprivacy release判断を混同しない。"
    },
    {
      "heading": "OPEN_UNKNOWN",
      "text": "- 2026-09-16時点の主要commercial synthetic-data provider各社について、plan/region/data-residency/training-retention/API revisionを横断固定していない。\n- SDV current exact release/tag/commitは将来update workで再pinする。\n- Text/LLM synthetic dataに対するattack suiteの標準化は成熟途上で、単一suiteを十分条件とはできない。"
    },
    {
      "heading": "STOP_CONDITION",
      "text": "Core採用判断に必要な概念分離、formal privacy boundary、implementation contract、utility/privacy/fairness/downstream evaluation、failure envelope、後続provider再確認項目まで揃ったためEA-080のbounded packetを閉じる。同一source version+claim relationは再採集せず、将来はcurrent provider/versionまたは新しいprivacy attack/standard差分のみ更新する。\n\n---"
    },
    {
      "heading": "AUDITED_CURRENTNESS_BOUNDARY",
      "text": "NIST SP 800-226 privacy-evaluation guidance remains distinct from synthetic generation, de-identification and attack-based evaluation."
    }
  ],
  "human_url": "/tools/ea-080-synthetic-data",
  "markdown_url": "/agent/article.md?path=%2Ftools%2Fea-080-synthetic-data",
  "json_url": "/agent/article.json?path=%2Ftools%2Fea-080-synthetic-data"
}