2026-02-20 -- Federated Analytics
- Federated analytics endpoints: Run descriptive stats, t-tests, ANOVA, and chi-square tests across federated data without centralizing it.
- Privacy budget tracking: Track cumulative epsilon spend per federation with history and cost estimation.
- Compositional DP accounting: Automatic privacy budget composition across multiple analytics queries.
2026-02-10 -- Experiments & A/B Testing
- Experiment framework: A/B test model versions with controlled traffic splits and device group targeting.
- Statistical analysis: Built-in significance testing for experiment results.
- Experiment analytics API: Retrieve per-variant metrics programmatically.
2026-01-28 -- Advanced Aggregation Strategies
- Krum and Multi-Krum: Byzantine-robust aggregation for adversarial environments.
- FedMedian and FedTrimmedAvg: Outlier-resistant aggregation without specifying attacker count.
- SCAFFOLD: Variance reduction with control variates for non-IID data.
2026-01-15 -- Compliance Presets
- HIPAA, GDPR, PCI, SOC 2 presets: One-command compliance configuration via
octomil init --compliance <preset>.
- Immutable audit logs: Configurable retention (up to 6 years for HIPAA).
- Compliance reports:
octomil org compliance-report --format pdf.
2026-01-05 -- Model Rollouts
- Progressive rollouts: Canary deployments with configurable percentage ramps.
- Automatic rollback: Rollback on error rate thresholds.
- Device group targeting: Roll out to specific device groups before fleet-wide deployment.
2025-12-18 -- Secure Aggregation
- Secure aggregation: Server cannot inspect individual client updates. Configurable reconstruction threshold.
- Differential privacy: Gradient clipping and calibrated noise injection with configurable epsilon/delta.
- Control plane: FastAPI server with device management, model registry, and training round orchestration.
- Python, iOS, Android SDKs: On-device training and inference.
- Dashboard: Device fleet monitoring, model versions, deployment management.
- FedAvg, FedProx, FedOpt, FedAdam: Core aggregation strategies.