Add technical architecture documentation for cognitive defence system

[Copilot]

Provides comprehensive technical writeup of the federated learning cognitive defence architecture for academic presentation, covering data distribution methodology, experiment execution, and defence strategy comparisons.

Created Document: TECHNICAL_ARCHITECTURE.md

24KB technical writeup (678 lines) structured as:

Core Architecture

• 4-layer system design (Orchestration → Server → Client → Defence/Attack)
• Component interactions and modular design principles

Data Distribution Strategy

• Non-IID partitioning via Dirichlet distribution: Dir(α × [1,...,1]) with α=0.5
• Heterogeneous client data reflecting realistic federated scenarios
• MNIST dataset: 60K train samples distributed across clients, 10K centralized test

Experiment Methodology

• FedAvg protocol with defence-enhanced aggregation
• Batch client spawning with resource monitoring
• YAML configuration system with deterministic seeding for reproducibility

Cognitive Defence (OODA + MAPE-K)

Four-phase cycle:

• Observe: Extract parameter norms and update magnitudes
• Orient: Z-score anomaly detection: Z_i = (||Δθ_i|| - μ) / σ, threshold at 2.0
• Decide: Reputation-based weighting with decay (0.8) on anomaly, reward (0.05) on normal
• Act: Weighted aggregation: θ_global = Σ(reputation_i × n_i × Δθ_i) / Σ(reputation_i × n_i)

Key differentiators: historical context (100-round window), gradual trust adjustment, explainable decisions with evidence

Strategy Comparison

Strategy	Approach	Complexity	Memory	Adaptation
Cognitive Defence	Adaptive weighted averaging	O(nd)	O(R)	Per-client, per-round
Krum	Selection-based (distance)	O(n²d)	O(1)	None
Trimmed Mean	Outlier removal (β trim)	O(nd log n)	O(1)	None
FedAvg	Simple weighted average	O(nd)	O(1)	None

Robustness analysis: Cognitive defence provides soft aggregation vs. Krum's binary selection and Trimmed Mean's fixed trimming. Maintains client identity tracking absent in stateless methods. Lower computational cost than Krum (O(nd) vs O(n²d)).

Attack Coverage

• Label flipping, gradient noise, model replacement simulations
• Evaluation metrics: centralized accuracy, reputation tracking, anomaly detection rates

Additional Documentation

• README.md updated with link to technical documentation
• Sections on implementation technologies, experimental workflow, research contributions

Original prompt

provide a small technical writeup of the architecture of this defence design. the write up should be presentable to an academic in the field to make him understand our research. it should focus on how the data is divided, how the experiments are run, how our cognitive defence strategy works, and how it differs from the krum, trimmed mean and fedavg strategies.

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.