Scholarlink

<p>AI-driven anomaly detectors in 5G renewable energy IoT and industrial systems lack unified governance: they operate opaquely, exhibit protocol-class bias, and expose training data to inference attacks. This paper presents the Ethical AI Governance Framework (EAGF), which maps four EU AI Act pillars, transparency (<i>C</i>), fairness (<i>RP</i>/<i>FPRP</i>), privacy (<i>P</i>), and accountability (<i>A</i>) to computable engineering metrics that are jointly governed within one training-and-deployment lifecycle: fairness and privacy are co-optimized via a Pareto-guided multi-objective procedure with domain-adaptive fairness loss selection, transparency is structurally controlled through clarity-triggered pruning, and accountability is audited post hoc, with all four scores aggregated into a composite Trust Index (TI). Evaluated across two domains: on a biometric task (10,021 images, ten seeds), EAGF raises TI by <InlineEquation ID="IEq1"><EquationSource Format="TEX">\(+38.97\%\)</EquationSource></InlineEquation> (<InlineEquation ID="IEq2"><EquationSource Format="TEX">\(0.565\rightarrow 0.785\)</EquationSource></InlineEquation>), improves recall parity by <InlineEquation ID="IEq3"><EquationSource Format="TEX">\(+15.1\%\)</EquationSource></InlineEquation>, and enhances privacy by <InlineEquation ID="IEq4"><EquationSource Format="TEX">\(+18.8\%\)</EquationSource></InlineEquation>; on the real-world Edge-IIoTset intrusion-detection benchmark (157,800 samples, five seeds), EAGF achieves <InlineEquation ID="IEq5"><EquationSource Format="TEX">\(+69.3\%\)</EquationSource></InlineEquation> TI gain (<InlineEquation ID="IEq6"><EquationSource Format="TEX">\(0.358\rightarrow 0.606\)</EquationSource></InlineEquation>) and <InlineEquation ID="IEq7"><EquationSource Format="TEX">\(+56.4\%\)</EquationSource></InlineEquation> FPR parity improvement, with only <InlineEquation ID="IEq8"><EquationSource Format="TEX">\(+0.2\)</EquationSource></InlineEquation>&#xa0;ms forward-pass inference overhead. Joint multi-pillar governance substantially outperforms model-level-only approaches across both domains; the accountability infrastructure contributes a large and explicitly quantified fraction of total TI gains, underscoring that governance readiness requires both algorithmic and operational investments.</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

EAGF: a four-pillar ethical AI governance framework for trustworthy cybersecurity in 5G renewable energy IoT systems

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月23日

查看全文
  • Salman Jan,
  • Ali Akarma,
  • Toqeer Ali Syed,
  • Munir Azam Muhammad,
  • Shahid Kamal

摘要

AI-driven anomaly detectors in 5G renewable energy IoT and industrial systems lack unified governance: they operate opaquely, exhibit protocol-class bias, and expose training data to inference attacks. This paper presents the Ethical AI Governance Framework (EAGF), which maps four EU AI Act pillars, transparency (C), fairness (RP/FPRP), privacy (P), and accountability (A) to computable engineering metrics that are jointly governed within one training-and-deployment lifecycle: fairness and privacy are co-optimized via a Pareto-guided multi-objective procedure with domain-adaptive fairness loss selection, transparency is structurally controlled through clarity-triggered pruning, and accountability is audited post hoc, with all four scores aggregated into a composite Trust Index (TI). Evaluated across two domains: on a biometric task (10,021 images, ten seeds), EAGF raises TI by \(+38.97\%\) (\(0.565\rightarrow 0.785\)), improves recall parity by \(+15.1\%\), and enhances privacy by \(+18.8\%\); on the real-world Edge-IIoTset intrusion-detection benchmark (157,800 samples, five seeds), EAGF achieves \(+69.3\%\) TI gain (\(0.358\rightarrow 0.606\)) and \(+56.4\%\) FPR parity improvement, with only \(+0.2\) ms forward-pass inference overhead. Joint multi-pillar governance substantially outperforms model-level-only approaches across both domains; the accountability infrastructure contributes a large and explicitly quantified fraction of total TI gains, underscoring that governance readiness requires both algorithmic and operational investments.