Skip to content

architecture.md

Latest commit

 

History

History
141 lines (100 loc) · 2.77 KB

architecture.md

File metadata and controls

141 lines (100 loc) · 2.77 KB

Architecture

System Overview

Production federated learning system with privacy, security, and communication efficiency.

Kubernetes Cluster
├── FL Server (SuperLink)
├── FL Clients (Workers)  
├── MLflow (Tracking)
└── Prometheus + Grafana (Monitoring)

Core Components

Data Loading

Thread-safe singleton pattern for dataset management.

class FederatedDataLoader:
    _lock = threading.Lock()
    _instances: Dict[str, 'FederatedDataLoader'] = {}

Features:

  • Lazy initialization
  • Patient-based partitioning
  • Configurable transforms
  • 80/20 train/test split

Gradient Compression

Hybrid compression combining quantization and sparsification.

Pipeline:

  1. Add error feedback from previous round
  2. Apply top-k sparsification (keep 10%)
  3. Quantize to 8 bits
  4. Store error for next round

Expected: 20-50x compression, <2% accuracy loss

Byzantine-Robust Aggregation

Defense mechanisms:

Multi-Krum: Select k clients with smallest pairwise distances

Trimmed Mean: Sort element-wise, trim extremes, average

Anomaly Detection: Z-score based outlier detection

Handles up to 30% malicious clients

Privacy Validation

Membership inference attacks for empirical privacy auditing.

Process:

  1. Train shadow models
  2. Extract prediction confidence
  3. Train attack model
  4. Evaluate on target

Metrics: Attack accuracy, AUC-ROC, privacy budget correlation

Training Round Flow

1. Server broadcasts model
2. Clients train locally
3. Clients compress gradients
4. Clients send updates
5. Server detects Byzantine clients
6. Server aggregates (robust)
7. Server updates model
8. Repeat

Deployment

Kubernetes Resources

Server: 1 replica, 512Mi-1Gi RAM, 0.5-1 CPU

Clients: 5-20 replicas (HPA), 1-2Gi RAM, 1-2 CPU

MLflow: StatefulSet, 10Gi storage

Monitoring: Prometheus + Grafana

Configuration

Helm chart with configurable:

  • Number of clients
  • Training rounds
  • Learning rate
  • Privacy budget
  • Compression settings

Security

  1. Differential Privacy (DP-SGD)
  2. Secure Aggregation (encrypted updates)
  3. Byzantine Robustness (Multi-Krum, Trimmed Mean)
  4. TLS/mTLS (encrypted communication)

Monitoring

Metrics:

  • Training: loss, accuracy, convergence
  • Communication: bytes, compression ratio, bandwidth
  • Security: Byzantine detection, privacy budget
  • System: CPU, memory, latency

Dashboards:

  1. Training Overview
  2. Communication Efficiency
  3. Byzantine Detection

Best Practices

Development:

  • Use virtual environments
  • Run tests before committing
  • Follow PEP 8
  • Add type hints

Deployment:

  • Use Helm
  • Set resource limits
  • Enable autoscaling
  • Configure health checks

Security:

  • Enable DP for sensitive data
  • Monitor Byzantine behavior
  • Rotate certificates
  • Audit privacy regularly