SkandaNet

Classical concepts for biometric security systems.

SkandaNet implements classical baselines that can be compared against the IndraNet Quantum SVM approach:

• Siamese CNN with hard negative mining, contrastive loss, supervised contrastive loss, ArcFace logits, and center loss
• Triplet Network with semi-hard negative mining, triplet margin loss, supervised contrastive loss, ArcFace logits, and center loss
• Discriminative Autoencoder with reconstruction, classification, supervised contrastive, and triplet losses
• RBF-SVM over the same normalized Iris measurements

The current offline benchmark uses sklearn.datasets.load_iris because this repository does not include a production biometric iris-image corpus. The CNN models receive deterministic image-like texture tensors derived from the Iris measurements, while RBF-SVM uses the standardized tabular measurements directly. Training uses a train/validation/test split, AdamW, learning-rate scheduling, early stopping, validation-selected model checkpoints, and multi-seed hyperparameter sweeps.

Run the benchmark

python scripts/benchmark_classical_models.py --epochs 60 --out-dir benchmarks

The default sweep tries embedding dimensions 8,16,32,64, learning rates 0.001,0.003,0.01, and seeds 2026,2027,2028.

Generated artifacts:

• benchmarks/classical_model_benchmark.json
• benchmarks/classical_model_accuracy.csv
• benchmarks/classical_model_biometric_metrics.csv
• benchmarks/classical_model_convergence.csv
• benchmarks/classical_model_convergence.svg

The biometric report includes ROC-AUC, FAR, FRR, EER, and TAR at 1% FAR.

Run tests

python -m pytest