Document Spoofing

Document spoofing is the presentation-side attack on document verification: rather than forging the document's content, the attacker manipulates how it is captured — photographing a screen displaying a document image, submitting a photo of a photocopy, replaying a capture recorded elsewhere, or injecting a prepared image directly into the capture stream, bypassing the camera entirely. The distinction from forgery matters operationally: a spoofed capture can carry a perfectly genuine document (someone else's, or one the presenter shouldn't control), and content-focused forensics will find nothing wrong with it.

Detection therefore reads the capture, not just the content. Screen-replay attacks betray themselves through moiré interference patterns, pixel grids, specular glare characteristic of displays, and unnatural color spectra; print-copy attacks through paper texture, halftone dots, and missing depth; and physical-document authenticity checks look for what only the real object exhibits — holographic elements shifting across frames, optically variable inks, embossing shadows — which is why serious ID verification uses short video or multi-frame capture with prompted movement rather than a single still. Injection attacks, the deepest variant, are fought below the image: attested capture SDKs, device integrity checks, and cryptographic binding of the capture session, because no pixel analysis catches an image that never passed through optics.

Spoofing defense matters most in remote onboarding and KYC, where the document check and its companion selfie/liveness check are the identity perimeter — and where generative AI has raised the stakes by making both fake documents and fake faces cheap. The architecture that holds is layered: capture-side controls (attested SDKs, liveness prompts), presentation-attack detection models on the received imagery, content forensics on the document itself, and cross-checks against external records — with the scoring fused, thresholds tuned to the channel's risk, and the adversarial reality accepted: this is a moving contest, and detection models age faster here than anywhere else in document AI.