Biometric Authentication Accuracy Using Siamese Neural Network in Online Exam Proctoring Systems
Annotatsiya
Online exam proctoring systems increasingly rely on automated facial verification to confirm that the individual taking a remote exam matches the enrolled student, a task structurally suited to one-shot or few-shot verification rather than closed-set classification, since the system must generalise to students never seen during model training. Siamese neural networks, which learn a discriminative embedding space via paired or triplet comparison rather than per-identity classification, are well matched to this requirement: verification reduces to a distance threshold in embedding space rather than retraining a classifier each time a new student enrolls. This paper applies a Siamese network (twin ResNet-18 towers, shared weights, 128-dimensional embedding, triplet loss with margin 0.4) to identity verification across a six-institution consortium of 64,000 enrolled students and 384,000 proctored verification events (64,000 × 6 mean exams per student), benchmarking against Eigenfaces/PCA, a closed-set softmax CNN classifier, and a pretrained FaceNet-style embedding without domain-specific fine-tuning. The fine-tuned Siamese model achieves an Equal Error Rate of 1.8% and AUC=0.984, substantially outperforming Eigenfaces (EER 8.4%; 78.6% relative reduction), the closed-set CNN (5.2%; 65.4%), and the non-fine-tuned pretrained embedding (3.8%; 52.6%). Critically, error rates are not uniform across the enrolled population: stratifying by Fitzpatrick skin-type grouping shows EER of 1.2% for Type I–II (lighter) skin tones, 1.6% for Type III–IV, and 3.8% for Type V–VI (darker) skin tones—a 216.7% relative disparity between the lightest and darkest groups, consistent with documented findings elsewhere in the facial-recognition fairness literature and a finding this paper treats as a primary result rather than a footnote. Deployed across one academic term, the verification system reduced live human-proctor identity-review burden by 81.6 percentage points (from manual review of all sessions to automated pre-screening of 18.4% requiring human follow-up) and flagged 284 sessions for suspected impersonation, of which 247 (87.0%) were confirmed upon human review. Webcam-resolution sensitivity, lighting-condition sensitivity, a measured 27.0% miss rate against simple photo-based presentation attacks absent dedicated liveness detection, and the skin-tone disparity above each bound the system's current scope to checkpoint identity-screening with mandatory human review of any failed or low-confidence match, rather than autonomous high-stakes determination.
Hali tarjima qilinmagan