Deepfake Detection System Problem: Deepfake accessibility

Deepfake Detection System

Problem: Deepfake accessibility creates real risks—misinformation, identity fraud, digital trust erosion. I built a production-grade detector: fast, accurate, explainable, CPU-deployable.

Approach: Transfer learning with ResNet-50 (25.5M params, pre-trained ImageNet-1K V2). Rather than training from scratch on limited data, I leveraged rich hierarchical features that transfer exceptionally well to detecting artifact patterns (blending boundaries, color mismatches, compression artifacts).

Key Decisions:

Face-Centric Preprocessing: OpenCV DNN crops faces to 224×224, concentrating attention on artifact-rich regions (eyes, mouth, jawlines).

Custom Classification Head: Bottleneck design (2048→512→1) with Focal Loss, label smoothing, progressive unfreezing—prevents overfitting on ~4,000 Ciplab images.

Explainable AI: Custom Grad-CAM generates attention heatmaps showing exactly where the model detected artifacts.

Real-World Robustness: Trained on JPEG compression, Gaussian noise, blur augmentations to handle compressed video feeds.

Results:

AUC-ROC: 0.9424 | Accuracy: 87.25% | Precision/Recall: 87.27%/84.81%

Inference: <100ms per face on CPU

Live on Hugging Face Spaces (no GPU)

Tech: PyTorch, OpenCV, Albumentations, Gradio. Images, videos, live webcam—one codebase.

Demonstrates: End-to-end AI engineering with deliberate trade-offs (efficiency over raw accuracy), training discipline (Focal Loss, Cosine Annealing), and production-first thinking (CPU compatibility, interpretability, real-world robustness).

Demo: https://huggingface.co/spaces/Shri04/deepfake-detector