Skip to content

System_Card.md

Latest commit

 

History

History
49 lines (39 loc) · 3.29 KB

System_Card.md

File metadata and controls

49 lines (39 loc) · 3.29 KB
Raw

System Card: FaceID Verification Engine

1. System Overview

FaceID is a facial verification system designed to determine if two facial images belong to the same identity. It uses deep representation learning to map facial features into a 128-dimensional embedding space where similarity is measured via cosine distance.

  • Model Architecture: FaceNet (InceptionResNetV1)
  • Framework: DeepFace / TensorFlow
  • Dataset (Training): Pre-trained on MS-Celeb-1M
  • Inference Pipeline: Preprocessing -> Feature Extraction -> Cosine Similarity -> Calibrated Confidence

2. Intended Use

  • Primary Use Case: Identity verification for secure access or identity confirmation (1:1 matching).
  • Intended Users: System administrators and developers integrating facial verification into applications.
  • Environment: Containerized (Docker) or local CLI/Web interface.

3. Out-of-Scope Uses

  • Surveillance: Not designed for real-time identification from large crowds (1:N search).
  • Demographic Classification: Not intended to predict age, gender, or ethnicity.
  • High-Stakes Legal Decisions: Should not be the sole basis for legal or judicial outcomes without human oversight.

4. Performance Metrics (Final Release)

Evaluated on the LFW (Labeled Faces in the Wild) validation subset.

  • Operating Threshold: 0.35 (Cosine Similarity)
  • Accuracy: 84.6%
  • F1-Score: 0.8254
  • Avg. Latency (CPU): ~0.47s per image (single-pair pipeline ~0.94s)

5. Limitations and Failure Modes

  • Lighting: Performance degrades in extreme low-light or high-contrast shadows.
  • Pose: High-angle profiles or partially occluded faces (masks, sunglasses) significantly increase the False Negative Rate.
  • Resolution: Low-resolution images ( < 64x64 pixels) result in unreliable embeddings.
  • Similarity in Identity: High similarity scores may occur between look-alikes or biological relatives (False Positives).

6. Fairness and Risks

  • Demographic Bias: The system was primarily evaluated on the LFW (Labeled Faces in the Wild) dataset. LFW is known to have a demographic skew:
    • Ethnicity: Predominantly Caucasian individuals (~70%+).
    • Gender: Significant bias towards male subjects (~75%).
    • Age: Most subjects are between 20-50 years old.
  • Risk Mitigation: Performance may be significantly lower for underrepresented groups (e.g., non-white ethnicities, children, or elderly individuals). Users should perform their own local fairness audits with balanced datasets before deployment in diverse environments.
  • Misuse Concerns: The system should not be used for mass surveillance or without the explicit consent of the subjects, except where legally permitted.
  • Privacy: The system processes sensitive biometric data. Developers must ensure compliance with data protection regulations (e.g., GDPR, BIPA).

7. Operational Constraints

  • Hardware: Minimum 4GB RAM required. CPU-only inference takes ~1.7s; GPU acceleration is recommended for batch processing.
  • Input Format: Supports standard formats (JPG, PNG). Images must contain a detectable human face.

8. Reproducibility

The final release is tagged as v1.0-final. Full instructions to reproduce results are available in the Reproducibility Checklist.