Skip to main content
Build with Ultralytics Platform·Deploy and Monitor·Lesson 9/10
Lessonintermediate

Privacy, Regions, and Compliance

Where your data lives, where compute happens, and what dataRegion means in practice.

Production computer vision often touches sensitive data — facial regions, license plates, manufacturing IP. "Just use the cloud" can violate a contract. Ultralytics Platform supports regional placement so data privacy and compute stay where they need to, with the legal posture spelled out in the Ultralytics privacy policy. Picking the right dataRegion early avoids expensive migrations later.

Outcome

Choose a dataRegion that meets your compliance constraints and understand the tradeoffs.

Fast Track
If you already know your way around, here's the short version.

  1. us / eu / ap map to compute and storage regions.

  2. Pick by where users / data live, not by latency alone.

  3. Source data, training data, and deployed compute should be in the same region.

  4. Cross-region transfer has compliance and cost implications — avoid when possible.

Hands-on

What dataRegion actually means

Ultralytics Platform onboarding region map with latency

Platform exposes three top-level regions:

RegionCompute / storage location
usUS-Central — close to North American users
euEU-West — close to European users (GDPR-friendly)
apAsia-Pacific — close to APAC users

dataRegion is set per workspace. Once set:

  • Datasets uploaded to the workspace stay in that region.
  • Training runs execute in that region.
  • Deployments are placed in that region by default.
  • Monitoring and logs live in that region.

When the choice matters

ConstraintRegion choice
EU customers, GDPReu
US federal customersus (FedRAMP-aligned data centers)
China usersap (Asia-Pacific footprint)
No specific constraintClosest to where your team works

If you have a contractual obligation about data residency, pick the matching region before uploading anything. Migrating later requires re-uploading, re-training, and re-deploying.

No silent cross-region transfer

Platform won't transparently move data between dataRegions. Every upload is sticky. This is the right behavior — silent transfer is a compliance landmine — but it means the default region matters more than usual.

Latency vs residency

The two pull in opposite directions:

  • Latency wants: compute in the region closest to your clients.
  • Residency wants: compute in the region matching your data law.

For most projects they align. For multinational projects they don't. Three patterns:

PatternTradeoff
Single global deployment in euHigher latency for non-EU users, simpler compliance
Per-region deployments, shared datasetNeed separate datasets per region, harder ops
Federated learning (one model, regional fine-tunes)Most complex; only worth it for very large fleets

For deeper formal guarantees, the differential privacy glossary entry is the right place to start; it complements rather than replaces region pinning.

Personal data and detection

Two specific concerns for CV:

  • Faces. If your detector returns face boxes (or implicitly, in cropped frames), you may be processing biometric data. Some jurisdictions require explicit user consent and additional safeguards.
  • License plates. Often classified as PII in EU and California. Detection of plates is generally OK; reading them (OCR) crosses a line in most jurisdictions.

The detector itself usually doesn't store the original image — but the request logs do, by default. Platform lets you disable image retention per deployment:

Deployment → Settings → Logs
[ ] retain request payloads (off by default in eu)

When in doubt, leave retention off and pull only the metadata (counts, confidence, latency).

Data minimization at the edge

For sensitive applications, do as much processing as possible on the edge before sending anything to the cloud:

  • Run inference on-device; send only metadata (counts, alerts) to the server.
  • Mask faces or plates before uploading frames for retraining — a small step that ties into broader AI ethics, AI safety, and algorithmic bias practice.
  • Aggregate per-camera-day instead of streaming raw frames.

This is also a cost optimization — uploading 100GB of video per camera per day is expensive even before compliance considerations.

A compliance audit checklist

Before deploying for a regulated customer, walk through:

  1. What data leaves the customer environment? (Source frames? Predictions? Both?)
  2. What region does it go to? (us, eu, ap?)
  3. Is request payload retention off?
  4. Does the customer's contract allow training on their data? (Often no — train on a separate dataset.)
  5. What's the data deletion process? (How quickly can you remove a customer's data on request?)
  6. Who has access? (Workspace permissions, audit logs.)

This list isn't exhaustive — get a real compliance person involved for any regulated deployment — but it catches the ones that bite first.

Try It

For your project, write down where the data is collected, who the users are, and what jurisdiction's law applies. Pick a dataRegion. If you can't pick clearly, write down the question — a compliance person can probably answer in five minutes.

Done When
You've finished the lesson when all of these are true.

  • You've chosen a dataRegion for your project.

  • Your workspace and deployments are in that region.

  • You know whether request payload retention should be on or off for your deployment.

What's next

Final lesson: the operational shape of a long-running CV project on Platform.