Himanshi Khanna

I've developed GeoDP at Tata Consultancy Services Research, a geometry-aware differential privacy framework for face biometric template protection. Using ArcFace embeddings and Fisher-guided feature analysis, I achieved 93.67% Rank-1 identification accuracy at ε = 12 on CASIA-WebFace while reducing reconstruction and membership-inference risk.

My work spans LLM evaluation and NLP systems. I improved Gemma 3 4B accuracy from 48% to 54.43% on the NTSE MCQ benchmark through prompt engineering, RAG, and self-consistency decoding, and built a persona-aware emotional support chatbot pipeline with Hugging Face Transformers, BERT, and Llama 2.

I've also worked on contrastive graph anomaly detection, lip-sync deepfake detection, and real-time object detection and tracking with YOLOv8 and OpenCV. My deepfake work improved validation accuracy from 60% to 70% and average precision from 0.40 to 0.60.

I'm pursuing a B.Tech in Data Science & AI at IIT Bhilai and enjoy turning machine learning research into measurable, practical systems.