LinkedIn · Email · 🤗 Live Demo
- 🤖 AI/ML Engineer · Computer Vision · Multimodal Systems
- 🔬 Building DeepTrace — deepfake detection at 0.9496 AUC using spatial + frequency + CLIP fusion
- 📄 Research paper under preparation for IEEE publication
- 🎓 B.Tech CSE (Data Science) · Graduating Jan 2027
- 🤝 Open to internships and research collaborations in CV / multimodal AI
- 📫 amoghsri02@gmail.com
Languages
AI / ML
Deployment & Tools
Most deepfake detectors collapse after social media compression — they over-rely on RGB features that JPEG strips away. DeepTrace fixes this.
| Branch | Method | What it catches |
|---|---|---|
| 🖼️ Spatial | EfficientNet-B0 | Texture, blending & geometry artifacts |
| 📡 Frequency | Block-wise DCT on YCrCb | Synthesis artifacts that survive compression |
| 🧠 Semantic | CLIP ViT-B/32 (frozen) | Generalizes to unseen manipulation techniques |
Three branches → cross-attention fusion → calibrated confidence + GradCAM heatmap
| Accuracy | ROC AUC | Decision Threshold |
|---|---|---|
| 0.90 | 0.9496 | 0.162 (temperature-scaled) |
→ Try the live demo · → View repo
- 🏅 Dean's List for Academic Excellence
- 💻 Active contributor in AI/CV projects and hackathons