DeepGuard: Deepfake Face Detection using ResNet + Frequency Analysis
Model Description
A deepfake detection model using a dual-stream architecture combining a spatial stream (ResNet50) and a frequency stream (DCT-based CNN) to identify AI-generated or face-swapped content in images and video.
Most deepfake detectors rely solely on spatial artifacts (blending edges, unnatural skin texture). This model additionally analyzes the frequency domain, where GAN-generated and diffusion-generated faces leave distinctive spectral fingerprints invisible to the human eye — making it substantially harder to fool with post-processing.
Architecture: Dual-Stream Fusion Network
Stream 1 — Spatial Stream (ResNet50)
- Backbone: ResNet50 pretrained on ImageNet
- Input: RGB face crop (224×224), normalized
- Modifications:
- Final FC layer replaced with 512-dim embedding layer
- Added CBAM (Convolutional Block Attention Module) after layer3 to focus on face boundaries, eyes, and mouth — regions most commonly artifacts in deepfakes
- Dropout(0.4) before embedding layer
- Output: 512-dim spatial feature vector
Stream 2 — Frequency Stream (DCT-CNN)
- Input: Grayscale face crop converted to frequency domain via
2D Discrete Cosine Transform (DCT)
- High-frequency components amplified (×3) before input — this is where GAN fingerprints concentrate
- Architecture: Lightweight 5-layer CNN
- Conv(1→32, 3×3) → BN → ReLU
- Conv(32→64, 3×3) → BN → ReLU → MaxPool
- Conv(64→128, 3×3) → BN → ReLU
- Conv(128→128, 3×3) → BN → ReLU → MaxPool
- GlobalAvgPool → FC(512)
- Output: 512-dim frequency feature vector
Fusion + Classifier
- Concatenate spatial + frequency vectors → 1024-dim
- FC(1024 → 256) → ReLU → Dropout(0.3)
- FC(256 → 1) → Sigmoid
- Output: Probability score [0,1] — 0 = real, 1 = fake
Full Inference Pipeline
Input: Image or Video │ ▼ ┌─────────────────────────┐ │ Face Detection │ RetinaFace or MTCNN │ │ Extract all faces → crop + align │ │ Skip frames with no detected face └─────────────────────────┘ │ ┌──┴──────────────┐ ▼ ▼ ┌──────────┐ ┌──────────────┐ │ RGB Crop │ │ Grayscale │ │ 224×224 │ │ → DCT │ │ Normalize│ │ → Amplify HF │ └──────────┘ └──────────────┘ │ │ ▼ ▼ ┌──────────┐ ┌──────────────┐ │ ResNet50 │ │ DCT-CNN │ │ + CBAM │ │ (5 layers) │ │ → 512-d │ │ → 512-d │ └──────────┘ └──────────────┘ │ │ └────────┬─────────┘ ▼ ┌─────────────────┐ │ Concatenate │ 1024-dim │ → FC(256) │ │ → Sigmoid │ └─────────────────┘ │ ▼ Per-face score: P(fake) [For Video]: Aggregate frame scores → temporal smoothing (rolling avg, window=8) → Final video-level verdict