ID Tech

Multimodal Biometrics (Fusion)

Combining two or more biometric signals (e.g., face + fingerprint) to boost accuracy, resilience, and spoof resistance.

Overview

Multimodal biometrics fuse signals (e.g., face + finger, iris + vein) to improve accuracy and resilience. Fusion can also harden systems against presentation attacks when combined with modality-specific PAD.

How it works

  • Feature-level fusion: combine feature vectors before matching.
  • Score-level fusion: normalize and combine matcher scores (e.g., sum, weighted, learned).
  • Decision-level fusion: combine accept/deny votes (e.g., AND/OR rules).

Common use cases

  • Border & national ID deduplication (1:N)
  • High-assurance workforce access
  • Financial KYC with PAD stacking

Strengths and limitations

Strengths: Higher accuracy; graceful degradation; spoof resistance.
Limitations: Cost/complexity; correlation between signals can cap gains; tuning/maintenance.

Key terms

  • Score fusion: Combining matcher scores, often after normalization.
  • Decision fusion: Using voting or logic rules on match outcomes.

References

Latest Data Cards

  • Data Card

    FBI Seeks Next-Gen Biometric Algorithms for NGI Upgrade

    2025-11-24CC-BY-4.0fingerprint-recognitionfacial-recognitioniris-recognitionmultimodal-biometrics

    An FBI RFI calls for vendors to provide next-generation fingerprint, face, iris, and tattoo algorithms to modernize the Next Generation Identification (NGI) system.

    • Request spans tenprint, latent, face, iris, and tattoo matching for large-scale identification workloads.
    • Vendors are expected to participate in NIST testing and supply performance data.
    • Signals continued NGI modernization and potential procurement roadmap.
See more data cards

Frequently Asked Questions

Which fusion levels are typical?
Sensor, feature, score, and decision-level fusion. Score-level is common in practice due to availability and interoperability.
What benefits should I expect?
Lower false rejects at fixed false accept rates; robustness to sensor/environment variability; better PAD via ‘liveness stacking’.
How do you tune thresholds across modalities?
Normalize scores per modality (e.g., z-norm) and set an operating point using development data to meet target FAR/FRR under fusion.