Insight
On-device vs cloud face recognition: when to choose an edge SDK
An edge face recognition SDK runs detection, liveness, and matching on the device itself. A cloud face recognition API sends an image or video frame to a vendor's servers and returns a result. Both are valid, and the choice is decided by four things: what happens when connectivity drops, what a verification costs at your volume, whether you need real liveness, and where biometric templates are allowed to live.
Choose an edge SDK when
- Your users are on shop floors, sites, vehicles, or rural locations where connectivity is intermittent.
- You verify the same people every day - attendance economics punish per-call pricing.
- You need liveness on every capture, not only on high-value transactions.
- Biometric data cannot leave the country, the campus, or the device for regulatory reasons.
- You cannot dictate handset quality, so the SDK has to run on entry-level Android hardware.
Side by side
| Dimension | Edge SDK | Cloud API |
|---|---|---|
| Where matching runs | On the phone, tablet, or on-premise gateway | Vendor's servers |
| Behaviour without internet | On-device 1:1, 1:N and passive liveness remain available | No function until connectivity returns |
| Latency | Sub-second, no round trip | Network-bound, varies by region and load |
| Cost per verification | Licensed, flat - suits high-volume daily attendance | Metered per API call - grows linearly with headcount |
| Liveness | Passive, video-based, on-device | Often active/challenge-based and separately priced; cheap image APIs have none |
| Data residency | Biometric templates never leave the device or site | Depends on vendor region availability |
| Best for | Attendance, field ops, elections, defence, regulated data | Low-volume, high-value transactions with reliable connectivity |
The liveness trap in cloud pricing
Cloud vendors sell genuine video liveness, but at KYC-grade pricing. At attendance volumes most teams quietly fall back to static image comparison instead, which has no liveness check - a printed photo or a screen replay passes. If your risk is proxy attendance, an SDK with passive, video-based liveness running on-device solves the problem at a cost you can actually run daily.