SynthLens β€” AI-Generated Image Detector

FDCS-Net V4 multi-branch deep learning classification and live forensic feature inspection

Drag & drop your image here, or browse

Supports JPG, JPEG, PNG, WEBP, BMP up to 15 MB

Sample Real Photograph Authentic Photo
sample_real.jpg Held-out Test Split Photo
Sample AI Generated Image AI Generated
sample_ai.jpg Synthetic Generator Test Image

ℹ️ Technical Details & Forensic Methodology β–Ύ

Architecture Overview: FDCS-Net V4

FDCS-Net V4 is a hybrid deep neural network engineered to identify synthetic imagery across disparate generative model architectures (Midjourney, DALLΒ·E, Stable Diffusion, GANs):

  1. Spatial Branch: Leverages an ImageNet-pretrained EfficientNet-B0 backbone to capture high-level semantic inconsistencies and unnatural physical textures.
  2. Color Stability Branch: Evaluates high-frequency perturbation resistance via depthwise convolutions on quantized color channels, exposing synthetic color distribution artifacts.
  3. Frequency Branch: Computes a log-scaled 2-D Fast Fourier Transform (FFT) magnitude spectrum to capture periodic grid fingerprints generated during transpose-convolution upsampling.
  4. Attention Fusion Layer: Branch feature embeddings are L2-normalized and passed to a learned softmax attention head to produce dynamically weighted representations.
  5. Leak-Free Decision Boundary: The classification threshold (0.80) was chosen strictly on the validation set using Youden's J statistic ($J = \text{TPR} - \text{FPR}$), ensuring completely unbiased held-out test set evaluation.
Input Image (256x256x3)
       β”‚
   β”Œβ”€β”€β”€β”΄β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
   β–Ό                   β–Ό                   β–Ό
Spatial Branch   Color Branch      Frequency Branch
EfficientNet-B0  Color Stability   2D-FFT Magnitude
(Pretrained)     Map β†’ CNN         Spectrum β†’ CNN
   β”‚                   β”‚                   β”‚
   └───► L2-Normalize each branch β—„β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                       β”‚
             Attention Fusion Head
               (Softmax Weights)
                       β”‚
              Classifier Head
             (Dense 128 β†’ Sigmoid)
                       β”‚
              Real / AI-Generated
                
⚠️
Forensic & Legal Disclaimer:
SynthLens produces probabilistic classifications based on learned statistical, frequency, and color-stability artifacts. It is designed for research, screening, and educational purposes and must not be considered definitive legal or judicial evidence.