Back to Engineering Projects
IEEE Published📚 AI Research
Efficient Deepfake Detection Using AI (IEEE)
An IEEE-published AI-driven media forensics framework for detecting manipulated facial content under heavy video compression and cross-platform re-encoding using deep learning techniques.
Benchmark Results & Metrics
Uncompressed Acc.
96.4%
Compressed Acc.
91.2%
Detection Latency
23ms/frame
Real-time FPS
42
01 // Overview
This IEEE-published study presents a media forensics pipeline that combines spatial and frequency domain features to isolate micro-level facial blends and compression-invariant artifact markers.
02 // The Problem
Web uploads (e.g. YouTube, WhatsApp) enforce lossy re-encoding which erases visual cues. Generic CNNs struggle to distinguish real videos from deepfakes under heavy compression.
03 // System Architecture
Spatial-Frequency Fusion Pipeline
Faces are extracted and processed through a dual-branch network: a spatial XceptionNet stream and an InceptionResNetV2 stream. Features are fused using cross-attention layers.
ARCHITECTURAL DATA FLOW:Compressed Video -> MTCNN Face Extract -> Dual Branch CNN -> Attention Merger -> Real vs Fake Class
Architecture Components:
MTCNN Face Alignment
Xception Spatial Encoder
InceptionResNet Spectral Branch
Attention Fusion Classification
04 // Implementation
Implemented in TensorFlow and trained on the FaceForensics++ corpus. We added custom data augmentations simulating compression at various CRF rates to ensure robustness.
05 // Execution Workflow
11. Extract frames from input video.
22. Crop and align faces using MTCNN.
33. Process faces through dual CNN streams.
44. Compute attention matrices to merge features.
55. Output deepfake probability and overlay visual bounding boxes.
06 // Technology Stack
Frameworks & Vision
TensorFlowKerasOpenCVMTCNN
Signal Processing
SciPy SpectrumNumPy
07 // Technical Challenges
⚠️Training models to remain robust when facing zero-day deepfake synthesis generators.
⚠️Overcoming GPU VRAM bottlenecks during batch training of deep dual-branch networks.
Lessons Learned
- •Analyzing frequency coefficients exposes generative upsampling patterns that spatial channels miss.
- •Applying aggressive adversarial compression during training boosts model generalization by 18%.
Future Improvements
- •Moving to temporal transformers (ViT) to trace cross-frame consistency anomalies.
- •Deploying quantized ONNX models to edge web extensions.