Security-camera AI architecture

On-device AI vs cloud AI in security cameras: privacy, speed and accuracy

“AI camera” says very little about where analysis happens. The camera, a hub or recorder, and a provider’s cloud can each perform different jobs—with different consequences for privacy, latency, bandwidth, subscriptions and offline protection.

Where can the intelligence run?

At the camera

On-device or edge AI

The camera’s processor analyses frames near the point of capture. It may identify people, vehicles, animals or line crossings before deciding whether to record or send an alert.

Inside the home

Hub, NVR or home-server AI

A local hub or recorder receives one or more streams and performs heavier analysis. This can give battery cameras more processing support or provide central search without sending every frame to the public cloud.

Provider servers

Cloud AI

Video, selected frames or event clips are uploaded for analysis. Large server models can support advanced search and frequent feature improvements, but depend more heavily on internet access, accounts and provider policy.

Split workload

Hybrid AI

The camera may detect basic motion or objects locally, then send only metadata or a chosen clip for richer cloud analysis. Many systems are hybrid even when marketing calls them “local” or “cloud.”

Do not confuse analysis with storage: a camera can analyse locally but upload recordings to the cloud, or record locally while sending snapshots to a cloud service for classification.

Direct comparison

Factor	On-device or local AI	Cloud AI	Hybrid design
Alert speed	Potentially very low latency because analysis is nearby.	Depends on upload, provider load and the return path.	Local first-stage alerts with cloud enrichment.
Internet outage	May continue if power, local networking and recording survive.	Analysis and remote alerts may stop.	Basic detection may remain while advanced features disappear.
Privacy exposure	Can reduce raw video leaving the property.	More data may be processed or stored off-site.	Depends exactly which frames, clips and metadata leave.
Model capability	Limited by the device’s processor, memory, heat and power budget.	Can use larger models and server resources.	Balances fast local filtering with heavier remote analysis.
Updates	Firmware/model updates may be slower or end with device support.	Provider can improve server models without replacing hardware.	Both firmware and cloud-service support matter.
Bandwidth	Can transmit only relevant events or metadata.	May require substantial and reliable upload capacity.	Usually lower than continuous cloud analysis, but verify.
Recurring cost	May work without a subscription; local storage still has costs.	Often tied to a paid plan and retention tier.	Core functions may be free while advanced search is paid.

Privacy: ask what leaves, not just where AI runs

Camera image → local detection → selected frame/clip or metadata → provider processing → storage → app access.
Ask about every arrow, not just the first box.

Data minimisation

Is full video uploaded or only event clips?
Does the cloud receive thumbnails, faces or labels?
Can cloud analysis be disabled separately from remote access?
Are privacy zones applied before upload?

Protection and access

Is data encrypted in transit and at rest?
Is end-to-end encryption available and what features does it limit?
Can staff or contractors access footage, and under what controls?
Are all viewers protected by strong authentication?

Retention and training

How long are video, thumbnails and event metadata retained?
Can each category be deleted and exported?
Is customer footage used to train or improve models?
Can that use be refused without losing core security?

Household and neighbours

Are face profiles created and where are they stored?
Does the view capture public space or neighbouring property?
Are microphone, recognition and sharing settings necessary?
Can roles restrict who sees indoor cameras?

Accuracy is a system result—not a single percentage

Laboratory accuracy does not predict performance at your doorway. Measure the complete chain: detection, classification, notification, human verification and response.

Failure	Typical causes	Useful test
Missed person	Backlighting, distance, blocked view, low resolution, battery sleep or narrow detection zone.	Walk every approach by day and night at different speeds and clothing.
False person alert	Shadows, reflections, statues, foliage, screens or partial objects.	Log alerts for a week and identify repeat triggers before reducing sensitivity.
Late alert	Weak Wi-Fi, slow upload, cloud latency, battery wake-up or notification delay.	Time from entering the zone to the alert on Wi-Fi and mobile data.
Wrong identity	Angle, lighting, masks, limited enrolment images or lookalikes.	Treat familiar-face labels as advisory; verify visually before acting.
No usable evidence	Compression, motion blur, poor angle, short clip or overwritten storage.	Export a real test clip and confirm faces, clothing and direction are usable.
Offline failure	Router, cloud, DNS, power or account outage.	Disconnect internet safely and document detection, recording and notification behaviour.

Safety rule: AI classification is decision support, not proof. Never confront a person or dismiss a smoke, CO, panic or intrusion warning solely because an AI label says an event is harmless.

Which architecture suits which job?

Prefer local-first

Fast perimeter actions

Local analysis is attractive for triggering a light, siren or recording immediately, especially where broadband is slow or unreliable.

Prefer cloud strength

Advanced search and remote simplicity

Cloud services may suit users who value natural-language search, off-site evidence, continually updated models and easy multi-location access.

Prefer hub/NVR

Several cameras and 24/7 recording

A capable local recorder can centralise storage and analysis, reduce repeated subscriptions and keep high-volume footage on site.

Prefer hybrid

Balanced household system

Local object detection and recording plus encrypted off-site event backup can retain resilience without uploading every second of every stream.

Buying checklist

Ask the manufacturer	Strong answer	Warning sign
Which detections run without internet or subscription?	Feature-by-feature list for the exact model and firmware.	“AI-powered” with no processing-location detail.
What is stored locally, on a hub and in the cloud?	Clear data-flow and retention explanation.	“Local privacy” while thumbnails or identities still leave unexplained.
What happens during internet loss?	Documented recording, detection and later synchronisation behaviour.	The app is the only way to discover failure.
How long are AI and security updates supplied?	Published support period and vulnerability-reporting route.	No commitment beyond the warranty.
Can clips and metadata be exported and deleted?	Simple owner-controlled export and deletion.	Account cancellation destroys data before useful export.
What requires payment?	Permanent free functions and paid additions listed separately.	Trial experience presented without explaining the post-trial camera.

Examples of different architectures

Apple states that HomeKit Secure Video is privately analysed and encrypted on a home hub before secure upload to iCloud. Axis describes edge analytics that can filter video locally and send relevant metadata for further server or cloud analysis. Eufy states that supported cameras perform specified detections on-device or through a local HomeBase. These examples demonstrate architecture choices, not blanket guarantees for every model or feature.

Official technical information checked 1 September 2026. Hardware generations, regions, subscriptions and firmware can change processing and storage behaviour.

Bottom line: choose on-device AI for low latency, reduced data movement and offline capability; cloud AI for server-scale features and easy off-site access; or a well-documented hybrid for both. The best choice is the one whose real data flow, failure behaviour and update policy match your risk—not the one with the longest AI feature list.

Check camera upload needs →Assess subscription value →Check vendor lock-in →