22 lines
1.3 KiB
Markdown
22 lines
1.3 KiB
Markdown
# Interplanetary Edge Orchestrator: Privacy-Preserving Federated Optimization
|
|
|
|
This repository contains a minimal, working Python simulation of a privacy-preserving
|
|
federated optimization layer designed for fleets of robotics operating with offline-first
|
|
connectivity in space habitats. It demonstrates a simple, DP-friendly aggregation of local
|
|
updates from multiple clients to form a global model.
|
|
|
|
Usage highlights:
|
|
- Lightweight Client and Server implemented in Python.
|
|
- Local data training using gradient descent for linear regression.
|
|
- Privacy-preserving flavor via optional noise on aggregated updates.
|
|
- Offline-first capability via local update caching (non-connected clients save updates to disk).
|
|
|
|
Privacy controls
|
|
- The system supports DP-friendly clipping of updates to bound sensitivity.
|
|
- Client.train accepts an optional clip_norm parameter (default None). If provided, per-update deltas are clipped to have L2 norm at most clip_norm before sending to the server.
|
|
- Server.aggregate also supports an optional clip_norm parameter to clip all incoming updates prior to averaging, providing an additional privacy safeguard.
|
|
- You can combine clipping with Gaussian noise (noise_scale) for stronger privacy guarantees.
|
|
|
|
How to run tests:
|
|
- This repository provides a test script via test.sh (see below).
|