Hello! I am an ML Engineer specializing in Computer Vision, Deep Learning, and AI Safety (Adversarial ML & Deepfake Detection). I build end-to-end pipelines — from robust data engineering to training and optimizing SOTA models for production.
- 🧠 Core Focus: Deepfake Detection, Adversarial Robustness & Video Analytics
- 🛠️ Engineering: Building automated preprocessing pipelines & production-grade evaluation code
- 📱 Optimization: Adapting heavy architectures for edge devices (MobileNetV4, ONNX Runtime)
- 📬 How to reach me: deepranse@gmail.com
- Languages: Python (Advanced), Bash
- Frameworks & Deep Learning: PyTorch, PyTorch Lightning, ONNX Runtime
- Computer Vision: OpenCV, MediaPipe, InsightFace, Vision Transformers (ViT)
- AI Safety & Adversarial ML: Adversarial Robustness Toolbox (ART)
- DevOps & MLOps: Docker, Git, Linux (Ubuntu/Debian)
deepfake-detection — R&D project for industrial deepfake detection on edge devices.
- Developed and evaluated two pipelines: heavy DINOv2 (22.1M params) and lightweight MobileNetV4 (9.2M params).
- Achieved 0.9988 ROC-AUC (98.30% Accuracy) with DINOv2 and 0.9932 ROC-AUC (96.64% Accuracy) with MobileNetV4.
- Proven that MobileNetV4 (64 MB) maintains competitive quality, making it ideal for edge inference.
- Built a custom dataset (43.7k+ videos) combining 7 benchmarks and custom AuthorDeepFake-6600 dataset generated via InSwapper & SDXL.
- Deployed on NVIDIA RTX 3090 for training and optimization.
adversarial_pipeline — Master's thesis on real-time adversarial attacks against face recognition.
- Adapted NI-FGSM and MI-FGSM optimization for live video streams (33.6 FPS on Apple M1 Pro via MediaPipe).
- Achieved 99.75% Attack Success Rate (ASR) against SOTA face recognition models with imperceptible perturbations (PSNR > 36 dB).
