Global consumer tech trains anti-spoofing on 8K+ recordings

Overview

A security camera that can't tell a face from a printed mask isn't a security camera.

A Global Consumer Tech Company building anti-spoofing for security and home monitoring needed multi-modal training data ” NIR, visible, and depth ” captured under controlled environmental conditions across real homes, not in a lab.

Firstsource delivered 8,000+ tri-camera recordings across 42 rental properties and multiple property types ” captured under NDA-compliant handling with zero data breach.

This was Intelligence that Operates: multi-modal anti-spoofing data captured under environmental rigidity, on a launch schedule.

Challenges

• Single-modality spoofing detection fails on the real-world attack surface. A printed mask fools an RGB-only model; a depth-only model fails on adversarial geometry. NIR + visible + depth fusion is the model the security camera needs.
• Environmental rigidity in rental properties is a discipline problem, not a hardware problem. Capturing under tight lighting and movement constraints across 42 uncontrolled environments requires onsite quality discipline that one-off recording protocols can't sustain.
• Anti-spoofing data is also prototype-IP data. Capturing on the client's hardware in third-party properties means the data pipeline has to protect both the recordings and the underlying prototype ” no leakage, no exposure.

How We Made It Happen

We ran multi-modal capture with environmental discipline and security-first data handling across 42 properties and multiple property types.

• Tri-camera capture per scenario: NIR, visible, and depth. Synchronized multi-modal recordings delivered as one fused dataset for anti-spoofing model training.
• Physical AI delivered with environmental control across 42 properties. Lighting, movement, and scenario variation captured under consistent onsite protocols ” not best-effort recording.
• Red Teaming via adversarial spoofing scenarios. Recordings spanned a range of physical obfuscation scenarios ” designed to harden the model against real-world tampering.

Conclusion

Security AI ships when it can see what's trying to fool it. Firstsource ran multi-modal anti-spoofing capture across 42 properties with full data discipline ” turning fraud-detection data collection into Intelligence that Operates.