Scholarlink

<p>Public institutions in regulated environments often cannot share deployment logs, creating a “data desert” that limits failure-prediction research. This study proposes a hybrid generative framework that synthesises deployment logs while maintaining empirical privacy safeguards. Using sixteen months of logs from a Ghanaian public service portal, we combine Conditional Tabular GANs for structured metrics with a context-aware Low-Rank Adapted large language model for textual logs to produce a synthetic “digital twin” of the original dataset. Utility is assessed with a chronological Training-on-Synthetic, Testing-on-Real protocol, in which an XGBoost classifier trained only on synthetic data attains an F1-score of 0.9276, comparable to the real-data baseline. Privacy is evaluated using Gower-based distance to closest record, membership-inference auditing, attribute-inference auditing, and schema-based log validation, yielding no exact record replication, a very low empirical singling-out risk under the evaluated threat model, and no hallucinated error codes. A paired bootstrap hypothesis test confirms that the observed F1 difference is not statistically significant (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(p = 0.084\)</EquationSource></InlineEquation>), supporting the interpretation that synthetic data serves as a viable substitute rather than a superior alternative. The findings indicate that hybrid generative modelling can reduce empirical disclosure risk under the evaluated threat model and can thereby support responsible cross-agency data sharing for AIOps in regulated public sectors. The framework is presented as a proof of concept grounded in a single institutional setting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Overcoming the data desert: generative AI techniques for synthesising anonymised deployment logs in regulated public sectors

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Johnson Nuviadenu,
  • Themba Masombuka,
  • Ernest Mnkandla,
  • Malusi Sibiya

摘要

Public institutions in regulated environments often cannot share deployment logs, creating a “data desert” that limits failure-prediction research. This study proposes a hybrid generative framework that synthesises deployment logs while maintaining empirical privacy safeguards. Using sixteen months of logs from a Ghanaian public service portal, we combine Conditional Tabular GANs for structured metrics with a context-aware Low-Rank Adapted large language model for textual logs to produce a synthetic “digital twin” of the original dataset. Utility is assessed with a chronological Training-on-Synthetic, Testing-on-Real protocol, in which an XGBoost classifier trained only on synthetic data attains an F1-score of 0.9276, comparable to the real-data baseline. Privacy is evaluated using Gower-based distance to closest record, membership-inference auditing, attribute-inference auditing, and schema-based log validation, yielding no exact record replication, a very low empirical singling-out risk under the evaluated threat model, and no hallucinated error codes. A paired bootstrap hypothesis test confirms that the observed F1 difference is not statistically significant (\(p = 0.084\)), supporting the interpretation that synthetic data serves as a viable substitute rather than a superior alternative. The findings indicate that hybrid generative modelling can reduce empirical disclosure risk under the evaluated threat model and can thereby support responsible cross-agency data sharing for AIOps in regulated public sectors. The framework is presented as a proof of concept grounded in a single institutional setting.