EfficientNet-B3 Image Forgery Detection

This repository contains a fine-tuned EfficientNet-B3 model for binary classification of synthetic identity-document images as genuine or forged.

The checkpoint is provided as a PyTorch state_dict:

efficientnet_b3_finetuned.pt

Model description

Architecture: EfficientNet-B3
Framework: PyTorch and timm
Task: binary image classification
Input: one identity-document image
Output: genuine or forged probability
Input resolution: 300 × 300 pixels

Labels:

0 = genuine / bona fide
1 = forged / tampered

Dataset

The model was trained and evaluated on the template/composite portion of the Synthetic dataset of ID and Travel Documents (SIDTD), generated using MIDV-2020 document templates.

Dataset composition used in this project:

Genuine images: 1,000
Forged images:  1,222
Total images:   2,222

The forged samples include the following SIDTD annotation types:

Crop_and_Replace
Inpaint_and_Rewrite

The dataset itself is not included in this repository. Please follow the SIDTD dataset terms and citation requirements before downloading or redistributing it.

Preprocessing

Each image is:

converted to RGB;
resized to 300 × 300 pixels;
converted to a tensor;
normalized using ImageNet mean and standard deviation:

mean = [0.485, 0.456, 0.406]
std  = [0.229, 0.224, 0.225]

Training used mild brightness, contrast, and saturation variation. Horizontal flipping was not used because mirrored identity documents are not realistic examples.

Training

The model was initialized from a pretrained EfficientNet-B3. Phase 1 trained the binary classifier head. The final checkpoint was then fine-tuned by unfreezing the final EfficientNet feature block and classifier head.

Fine-tuning used:

optimizer: AdamW;
learning rate: 1e-4;
weight decay: 1e-4;
epochs: 20;
binary cross-entropy equivalent: two-class cross-entropy loss;
deterministic stratified hold-out split: 80% train, 10% validation, 10% test.

Evaluation

On the fixed 223-image test split at a decision threshold of 0.5:

Accuracy:  98.21%
Precision: 97.60%
Recall:    99.19%
F1:        98.39%
ROC-AUC:   99.98%

Confusion matrix, with rows representing actual labels and columns representing predictions:

[[97,   3],
 [ 1, 122]]

Subtype recall on the forged test samples:

Crop-and-replace: 94.74%
Inpainting:       100.00%

These results are specific to the project’s synthetic dataset and split. They should not be interpreted as real-world identity-verification performance.

Usage

Install the required packages:

pip install torch torchvision timm pillow

Load the checkpoint:

from pathlib import Path

import timm
import torch
from PIL import Image
from torchvision import transforms


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = timm.create_model(
    "efficientnet_b3",
    pretrained=False,
    num_classes=2,
)
model.load_state_dict(
    torch.load(
        "efficientnet_b3_finetuned.pt",
        map_location="cpu",
        weights_only=True,
    )
)
model.to(device)
model.eval()

preprocess = transforms.Compose([
    transforms.Resize((300, 300)),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=(0.485, 0.456, 0.406),
        std=(0.229, 0.224, 0.225),
    ),
])

image = Image.open("document.jpg").convert("RGB")
input_tensor = preprocess(image).unsqueeze(0).to(device)

with torch.no_grad():
    probabilities = torch.softmax(model(input_tensor), dim=1)[0]

genuine_probability = float(probabilities[0])
forged_probability = float(probabilities[1])
label = "forged" if forged_probability >= 0.5 else "genuine"

print({
    "label": label,
    "genuine_probability": genuine_probability,
    "forged_probability": forged_probability,
})

Explainability

The project uses Grad-CAM with the final spatial convolutional layer of EfficientNet-B3. The resulting heatmap indicates image regions that contributed to the selected prediction.

Grad-CAM is an explanation aid and should not be interpreted as a guaranteed localization of the tampered region.

Intended use

This model is intended for:

educational computer-vision projects;
research prototypes;
synthetic identity-document forgery experiments;
demonstrations of explainable image classification;
offline error analysis.

Limitations

The training data is synthetic.
The model has not been validated on real identity documents.
Performance may depend on document templates, image quality, and dataset artifacts.
The binary output is not equivalent to full identity verification.
A positive or negative result should not be used as the sole basis for a KYC or access-control decision.
The reported test split is an image-level random split and may not measure generalization to unseen document families as strictly as a grouped or external split.

Citation and attribution

This model uses the SIDTD dataset and MIDV-2020-derived document templates. Please cite and follow the terms of the original SIDTD and MIDV-2020 resources when using this model or reproducing the experiments.

SIDTD project: https://github.com/Oriolrt/SIDTD_Dataset

License

The model repository is released under the MIT license. Dataset licensing, attribution, and redistribution terms remain subject to the original SIDTD and MIDV-2020 sources.

Downloads last month: -; Downloads are not tracked for this model. How to track