Privacy, Fairness and Responsible Use

Sensitive by default

Vision systems often capture people: faces, licence plates, homes, workplaces. Face recognition and biometric identification raise serious privacy and civil-liberty concerns and are restricted or regulated in many places. Vision models have shown performance gaps across skin tones, ages and genders when training data is unbalanced. Responsible practice: collect only what you need, blur or discard faces and plates when identity is not required, obtain consent and follow the law, evaluate performance per demographic group where relevant, and keep humans accountable for consequential decisions.

Privacy-by-design choices

Questions to settle before collecting images.

do we need identities, or only counts / presence?          -> prefer counts
can faces and plates be blurred at capture time?             -> blur on device
how long are raw images kept, and who can access them?       -> short retention, access logs
is there consent / legal basis and visible notice?           -> document it
has accuracy been checked across groups of people?           -> per-group evaluation
who reviews decisions that affect individuals?               -> named human owner

Process on the device when possible

Running models at the edge and sending only results (counts, alerts) avoids moving sensitive images at all.

त्वरित जाँच: Which practice reduces privacy risk in a people-counting system?

• Sharing images widely for debugging
• Storing all raw video forever
• Counting on the device and discarding or blurring faces
• Adding face recognition by default