EAGF: Ethical AI Governance Framework

A Four-Pillar Framework for Trustworthy Cybersecurity in 5G Renewable Energy IoT Systems

---

EAGF jointly optimises four governance pillars — transparency, fairness, privacy, and accountability — within a single training-and-deployment lifecycle. All four scores are aggregated into a composite Trust Index that quantifies end-to-end governance readiness. The framework is evaluated on two real-world cybersecurity domains: biometric access control and IIoT intrusion detection.

---

Results

Case Study 1 — Biometric Access Control

Dataset: EFR (10,021 images, 158 classes)  |  Model: ResNet-50 (ImageNet, GroupNorm)  |  Seeds: 10  |  Fairness metric: Recall Parity

Metric	Baseline (M0)	EAGF (M2)	Δ
Accuracy	84.63 ± 1.03 %	78.63 ± 2.52 %	−6.00 pp
Recall Parity	0.786 ± 0.021	0.905 ± 0.012	+0.119
Clarity	0.929 ± 0.028	0.961 ± 0.021	+0.032
Privacy	0.243 ± 0.009	0.289 ± 0.012	+0.046
Accountability	0.300 ± 0.000	0.983 ± 0.000	+0.683
Trust Index	0.565 ± 0.007	0.785 ± 0.008	+38.9 %

_{Exact Wilcoxon signed-rank test, p = 0.002 for all pairwise comparisons.}

Case Study 2 — IIoT Intrusion Detection

Dataset: Edge-IIoTset (157,800 flows, 1 : 5.8 imbalance)  |  Model: 2-layer BiLSTM (hidden = 128)  |  Seeds: 5  |  Fairness metric: FPR Parity

Metric	Baseline (M0)	EAGF (M2)	Δ
Accuracy (macro)	0.648 ± 0.025	0.665 ± 0.008	+2.6 %
FPR Parity	0.493 ± 0.085	0.771 ± 0.057	+56.4 %
Clarity	0.692 ± 0.043	0.739 ± 0.055	+6.8 %
Privacy	0.248 ± 0.003	0.248 ± 0.003	preserved
Accountability	0.000 ± 0.000	0.667 ± 0.000	+0.667
Trust Index	0.358 ± 0.013	0.606 ± 0.011	+69.3 %

_{Paired t-test, t(4) = 32.6, p < 0.001. System overhead: +0.2 ms latency, +5.8 MB memory.}

Highlights

• No pillar is redundant — leave-one-out ablation confirms every mechanism degrades TI when removed; accountability contributes the largest single share.
• Privacy holds — DP-SGD (ε = 3) reduces MIA AUC from 0.612 to 0.541 (CS1); near-random in CS2.
• Edge-deployable — +0.2 ms latency, +5.8 MB memory; SHAP runs asynchronously (~12 ms per batch).

---

Getting Started

1. Install

git clone https://github.com/aliakarma/eagf.git
cd eagf
python -m venv .venv && .venv\Scripts\activate   # Windows
# source .venv/bin/activate                       # Linux / macOS
pip install -r requirements.txt

2. Prepare Data

Case Study	Dataset	Instructions
CS1 — Biometric	EFR (Kaggle)	Place preprocessed arrays in data/biometric/efr_processed/. Use --demo for synthetic fallback.
CS2 — IIoT	Edge-IIoTset	Download CSV (~78 MB) and place at data/real_iot/edge_iiot.csv.

See data/README.md for full preprocessing instructions.

3. Run

Quick validation (single seed):

# CS1 — Biometric (demo data)
python run_eagf.py --config configs/biometric_default.yaml --seeds 42 --demo

# CS2 — IIoT
python run_full_pipeline.py --real_dataset edge_iiot --config configs/reiot_real.yaml --seeds 42

Full experiments (publication results):

# CS1 — 10 seeds
python run_eagf.py --config configs/biometric_default.yaml \
  --seeds 42 43 44 45 46 47 48 49 50 51

# CS2 — 5 seeds
python run_full_pipeline.py --real_dataset edge_iiot \
  --config configs/reiot_real.yaml --seeds 42 43 44 45 46

Both pipelines produce per-seed results, aggregate statistics (mean ± std, 95 % CI), ablation tables, and Pareto front visualisation (5 × 5 grid, 25 runs) under results/.

---

Method

Trust Index

The composite governance metric aggregates all four pillars with equal weight:

TI  =  0.25  ×  ( Clarity  +  Fairness  +  Privacy  +  Accountability )

Pillar	What it measures	CS1 metric	CS2 metric
Transparency	Explanation faithfulness	SHAP fidelity (Eq. 6-7)	SHAP fidelity
Fairness	Group-level parity	Recall Parity (ratio)	FPR Parity (ratio)
Privacy	Resistance to inference attacks	β e-ε + (1-β)(1-MIA)	Same
Accountability	Audit trail completeness	RSA-signed logs + compliance	Same

Architectures

	CS1 — Biometric	CS2 — IIoT
Model	ResNet-50 (ImageNet-pretrained, GroupNorm for Opacus, Dropout 0.3)	2-layer Bidirectional LSTM (hidden = 128, LayerNorm head)
Training	50 epochs, batch 32, cosine LR + 5-epoch warmup	30 epochs, batch 256, cosine LR + 3-epoch warmup
DP-SGD	ε = 3.0, max grad norm = 1.0	ε = 3.0, max grad norm = 1.0
Convergence	Early stop: ΔTI < 0.002 over 5 epochs	Same

---

Reproducibility

	CS1	CS2
Seeds	42 – 51 (10 runs)	42 – 46 (5 runs)
Statistical test	Exact Wilcoxon signed-rank (p = 0.002 floor)	Paired t-test, t(4) = 32.6
Determinism	NumPy + PyTorch seeds fixed per run	Same
Early stopping	ΔTI < 0.002 over 5 epochs or ε-budget exhausted	Same

---

Repository Structure

eagf/
├── run_eagf.py                      # CS1 biometric pipeline
├── run_full_pipeline.py             # CS2 Edge-IIoTset pipeline
│
├── src/
│   ├── models/
│   │   ├── __init__.py              # Model factory  (build_model)
│   │   ├── resnet50.py              # ResNet-50 + GroupNorm
│   │   ├── lstm.py                  # Bidirectional LSTM
│   │   └── tabular_mlp.py           # Tabular MLP fallback
│   │
│   ├── training/
│   │   ├── eagf_trainer.py          # Governance-aware training loop
│   │   ├── fairness_loss.py         # Differentiable RP / FPRP loss
│   │   └── pareto_trainer.py        # 5×5 Pareto front exploration
│   │
│   ├── evaluation/
│   │   ├── ablation.py              # 6-variant + leave-one-out ablation
│   │   ├── statistics.py            # Wilcoxon & paired t-test
│   │   ├── mia_attack.py            # Yeom loss-threshold MIA
│   │   ├── audit_logger.py          # RSA-SHA256 cryptographic audit
│   │   └── report_generator.py      # Aggregate report generation
│   │
│   ├── metrics/
│   │   ├── clarity.py               # SHAP-based clarity score
│   │   ├── fairness.py              # Recall Parity, FPR Parity
│   │   ├── privacy.py               # DP accounting + MIA score
│   │   ├── accountability.py        # Audit + trace + compliance
│   │   └── trust_index.py           # Composite Trust Index
│   │
│   ├── utils/
│   │   ├── data_loader.py           # Dataset dispatcher
│   │   ├── biometric_pipeline.py    # Image pipeline (CS1)
│   │   ├── edge_iiot_loader.py      # Edge-IIoTset loader (CS2)
│   │   ├── preprocessing.py         # Normalisation & DP noise
│   │   └── visualisation.py         # Figures
│   │
│   └── baselines/
│       ├── aif360_dp_pipeline.py    # AIF360 reweighing
│       ├── fairlearn_baseline.py    # Fairlearn ExponentiatedGradient
│       └── joint_dp_fair_baseline.py
│
├── configs/
│   ├── biometric_default.yaml       # CS1 config
│   ├── reiot_real.yaml              # CS2 config
│   └── compliance_checklist*.yaml   # Compliance templates
│
├── docs/
│   └── EAGF.tex                     # Paper source
│
├── data/                            # Dataset storage (not committed)
├── results/                         # Experiment outputs
└── figures/                         # Generated visualisations

---

Citation

If you use this code in your research, please cite:

@article{jan2026eagf,
  title   = {{EAGF}: A Four-Pillar Ethical {AI} Governance Framework for
             Trustworthy Cybersecurity in {5G} Renewable Energy {IoT} Systems},
  author  = {Jan, Salman and Syed, Toqeer Ali and Akarma, Ali and
             Muhammad, Munir Azam and Kamal, Shahid},
  year    = {2026},
  journal = {Under review}
}

---

Documentation  ·  Data Setup  ·  Configs  ·  License

MIT License © 2026 Ali Akarma