Privacy-Preserving Visual Localization

VPS knows where you are. That's its job. But it shouldn't know more than necessary, and that information shouldn't be exploitable.

Privacy Threats

Location History

Sequential VPS queries reveal movement patterns:

• Where you go
• When you go there
• How long you stay

Image Content

Images sent for localization could reveal:

• People you're with
• Activities you're doing
• Private spaces

Aggregation

Even anonymous data, aggregated, reveals patterns:

• Popular locations
• Traffic flows
• Gathering patterns

Defense: On-Device Processing

First principle: images never leave the device.

Camera → On-Device Feature Extraction → Features → Cloud
         (full image here)              (no image content)

Features are designed to be non-invertible:

• Can match to map, but can't reconstruct original image
• Tested with adversarial image reconstruction attacks
• Acceptable leakage: general scene type (indoor/outdoor), not specific content

Defense: Minimal Collection

Only collect what's needed:

• Features, not images
• Approximate location, not precise trajectory
• Session ID, not user ID

Retention policy:

• Query features deleted immediately after localization
• Aggregate statistics retained for service improvement
• No personal location history stored

Defense: Differential Privacy

Add calibrated noise to prevent individual identification:

For location aggregates:

True count + Noise(scale=ε) = Published count

Guarantees: can't determine if any individual was present.

For feature matching:

True match score + Noise = Reported score

Prevents precise timing inference while maintaining localization function.

Defense: User Control

Users must have meaningful control:

• Clear opt-in for VPS features
• Easy opt-out without losing device function
• Data download (see what we have)
• Data deletion (remove all traces)

Auditing and Verification

How do users trust our claims?

• Regular third-party privacy audits
• Open-source privacy components (where possible)
• Bug bounty for privacy vulnerabilities
• Transparency reports on data requests

The Performance Trade-off

Privacy measures have costs:

• On-device processing uses more battery
• Differential privacy reduces accuracy
• Limited collection prevents some features

We accept these costs. Privacy is a feature, not a bug.

Regulatory Landscape

GDPR, CCPA, and emerging regulations shape our design:

• Data minimization (collect minimum necessary)
• Purpose limitation (use only for stated purpose)
• Storage limitation (delete when not needed)
• User rights (access, correction, deletion)

Building compliance in, not bolting it on.