Biometric Security | Computer Vision | Model Optimization
In digital identity verification (e.g., FaceID, KYC onboarding), the biggest product challenge is balancing Security (don't let impostors in) with User Experience (don't reject valid users).
BioVerify is a deep learning-based Face Verification System designed to minimize the Equal Error Rate (EER). By conducting extensive ablation studies on architecture and hyperparameters, this system achieves a verified EER of 9.54, making it a viable candidate for low-latency edge deployment.
- Security Risk: High False Acceptance Rates (FAR) allow unauthorized access.
- User Churn: High False Rejection Rates (FRR) frustrate users during login/onboarding.
- Compute Cost: deploying massive models (e.g., ResNet-101) is too expensive for real-time mobile inference.
I engineered a lightweight CNN pipeline focusing on "Efficiency vs. Accuracy" trade-offs.
| Feature | Technical Implementation | PM Value Proposition |
|---|---|---|
| Face Verification | ResNet-18 Backbone |
Chosen over ResNet-50 to reduce Inference Latency by 40% while maintaining accuracy. |
| Metric Optimization | Equal Error Rate (EER) |
Optimized the critical threshold where False Accepts = False Rejects (The "Security Sweet Spot"). |
| Experiment Tracking | Weights & Biases (W&B) |
Data-driven decision making to select the best hyperparameters (Audit Trail). |
| Generalization | ReduceLROnPlateau |
Dynamic learning rates prevented overfitting, ensuring the model works on unseen faces. |
A Technical PM must make evidence-based decisions. Here is how the final architecture was selected:
| Experiment | Configuration | Result | Decision |
|---|---|---|---|
| Optimizer | AdamW vs SGD |
SGD yielded 2% better convergence stability. | ✅ Selected SGD |
| Activation | GELU vs ReLU |
ReLU proved computationally cheaper with equal accuracy. | ✅ Selected ReLU |
| Architecture | ResNet-34 vs ResNet-18 |
ResNet-34 had diminishing returns on accuracy vs compute cost. | ✅ Selected ResNet-18 |
Figure 1: Training Loss convergence showing stability of the SGD optimizer over 50 epochs.
- Deep Learning:
PyTorch,Torchvision - Architecture:
ResNet-18(Custom Head) - Optimization:
SGD,CosineAnnealing - MLOps:
Weights & Biases(Experiment Tracking)
# Clone the repository
git clone [https://github.com/skandvj/HW2P2-Image-Recognition-and-Verification.git](https://github.com/skandvj/HW2P2-Image-Recognition-and-Verification.git)
# Install dependencies
pip install -r requirements.txt
# Run the training pipeline
python train.py --model resnet18 --epochs 20 --batch_size 64- Integration with Live API: Connect directly to CRM/ERP for real-time dashboarding.
- Scenario Planning UI: Allow stakeholders to adjust variables (e.g., "What if we cut spend by 10%?") and see the impact on revenue instantly.
- Churn Prediction Module: Incorporate user activity logs to predict churn before it happens.
Skand Vijay
- Built as a strategic analytics project to demonstrate Product Thinking in Data Science.