Private ML. Collaborative Intelligence.
Promise: From baseline to audit in under an hour. Epsilon budgeted. RBAC enforced. Reports ready.
Aegis is the open-source, compliance‑first orchestration layer for privacy‑preserving ML. It puts Differential Privacy (DP) and Federated Learning (FL) to work in real production—API‑first, audited, observable, and deployable on Docker/Kubernetes.
Ship private models faster—with hard privacy guarantees, enterprise controls, and audit‑ready reports. No vendor lock‑in. Python native. Batteries included.
Most FL/DP tools give you parts. Aegis gives you the whole product.
- Others give you a library; we give you a platform.
- Not just DP primitives or an FL loop, but an API, CLI, security controls, dashboards, and compliance reporting—wired together.
- Others make privacy optional; we make it operational.
- Epsilon targeting, step‑wise accounting, and budget guardrails are first‑class—not a README footnote.
- Others lock you in; we set you free.
- Open‑source, Flower‑based orchestration, Opacus DP, and standard Docker/Kubernetes deployments.
- Others bolt on compliance; we bake it in.
- Tamper‑evident audit logs, RBAC, mTLS, and one‑click Markdown/PDF reports you can hand to regulators.
This is the difference between a toolkit and a product. Between a demo and a deployment.
- Differential Privacy you can govern
- DP‑SGD via Opacus, epsilon targeting and step‑wise accounting; clipping, noise, sample rate, and delta as first‑class parameters.
- Guardrails and validation for privacy budgets; epsilon consumption surfaced in dashboards.
- Federated learning that’s robust, not fragile
- Flower orchestration with selectable robust aggregators (Krum, Trimmed Mean), participant authentication, straggler policy, and basic health.
- Security and compliance built‑in
- mTLS, RBAC (admin/operator/viewer), structured audit logs with hash chaining.
- One‑click compliance reports (Markdown/PDF) with GDPR/HIPAA/EU AI Act mappings and DPIA‑style risk notes.
- Production ergonomics
- FastAPI with OpenAPI schema, CLI, and industry playbooks (HIPAA/PCI/GDPR) to run full workflows without touching code.
- Prometheus/Grafana prewired; warmup script for lively dashboards; Docker/Kubernetes deploy artifacts.
- Start the stack
make docs-up- Run a playbook (HIPAA/PCI/GDPR)
bash docs/playbooks/healthcare_hipaa.sh- Watch dashboards and produce a report
# Grafana
open http://localhost:3000
# Status (viewer role)
curl -fsS -H 'X-Role: viewer' 'http://localhost:8000/training/status?session_id=hipaa_run' | jq .
# Compliance report (Markdown)
curl -fsS -H 'X-Role: viewer' 'http://localhost:8000/compliance/report' | jq -r .markdown > report.mdTeardown:
make docs-downPull the published image from GitHub Container Registry. Images are multi-arch (linux/amd64, linux/arm64).
# Latest stable
docker pull ghcr.io/clyrai/aegis:latest
# Or pin to a version
docker pull ghcr.io/clyrai/aegis:v0.1.2
# Run API on :8000
docker run --rm -p 8000:8000 ghcr.io/clyrai/aegis:v0.1.2Apple Silicon (M1/M2) note: Docker selects arm64 automatically. To force:
docker pull --platform linux/arm64 ghcr.io/clyrai/aegis:v0.1.2Verify platforms in the manifest list:
docker buildx imagetools inspect ghcr.io/clyrai/aegis:v0.1.2Optional: verify Cosign signature and provenance once public or when authenticated.
- Privacy Engine (
aegis/privacy_engine.py): Opacus DP‑SGD, RDP accounting, epsilon targeting. - Federated Coordinator (
aegis/federated_coordinator.py): Flower strategy; Krum & Trimmed Mean; auth & retries. - API (
aegis/api.py): FastAPI; mTLS, RBAC, audit logs; OpenAPI schema. - Compliance Reports (
aegis/compliance/report.py): Markdown/PDF; versions captured. - Deploy (
deploy/): Docker Compose, Kubernetes, Helm; secrets and healthchecks.
- Prometheus/Grafana provisioned; dashboards show RPS, latency, errors, epsilon consumption, scrape health.
- Security tests include membership inference/model inversion; RBAC & mTLS are validated in tests and examples.
- Healthcare collaboratives (HIPAA)
- Financial institutions (PCI)
- Retailers & advertisers (GDPR)
- Open, portable stack vs. vendor lock‑in; production features missing in raw FL/DP libraries.
- Compliance and audit “baked in,” not bolted on.
- Works with PyTorch/scikit‑learn/TensorFlow examples; bring your own models.
- Flower — excellent orchestration (we build on it), but no integrated DP accounting, RBAC/mTLS, or compliance reports.
- NVIDIA FLARE / OpenFL / FATE — strong for secure aggregation/MPC, heavier to operate; Aegis prioritizes DP governance + production ergonomics.
- FedML (SaaS) — managed convenience, but with vendor lock‑in; Aegis is open, on‑prem friendly.
- DP libraries (Opacus, TF‑Privacy, OpenDP) — great building blocks; Aegis is the production wrapper that unifies them.
- Optional secure aggregation modules; additional robust aggregation strategies
- Deeper policy presets for GDPR/HIPAA/PCI
- Extended attack simulations and reporting
We welcome issues and PRs. Please run:
ruff check && pytest -q && python -m mkdocs buildBuilt and maintained by clyrai.
Licensed under the Apache License, Version 2.0 (Apache-2.0). See LICENSE for details.