Sowaiba01/deepguard-ai

DeepGuard AI — EfficientNet-B4 Deepfake Detector

DeepGuard AI is a binary deepfake detection model fine-tuned on a custom face-swap dataset using EfficientNet-B4. It detects AI-generated face swaps (inswapper, SimSwap, DeepFaceLab, StyleGAN2) and provides GradCAM explainability heatmaps to highlight suspicious facial regions.

This repository also contains inswapper_128.onnx (InsightFace) used for adversarial face-swap generation in the research pipeline.

---

Model Files

File	Size	Description
efficientnet_b4_deepguard.pth	74.9 MB	v1 checkpoint — EfficientNet-B4 binary classifier
efficientnet_b4_deepguard_v2.pth	74.9 MB	v2 checkpoint (recommended) — fine-tuned on extended dataset
inswapper_128.onnx	554 MB	InsightFace face-swap model for generation pipeline

---

Performance (v2 — Validation Set)

Metric	Value
Accuracy	91.54%
Precision	91.73%
Recall	91.38%
F1 Score	0.9156
AUC-ROC	0.9486
False Positive Rate	1.27%

Cross-Method Benchmark

Method	Accuracy	AUC
StyleGAN2	94.8%	0.9820
inswapper_128	91.5%	0.9486
FaceSwap	89.3%	0.9380
DeepFaceLab	85.1%	0.9210
Stable Diffusion	78.4%	0.8940

---

Training Configuration

Parameter	Value
Base model	EfficientNet-B4 (timm)
Pre-training	ImageNet-21k
Input size	380 × 380
Training images	10,852
Optimizer	AdamW (lr = 3e-5)
Scheduler	Cosine annealing
Batch size	16
Epochs	8 (early-stop patience 3)
Augmentation	Flip, Rotate, Color jitter
Loss	Binary cross-entropy

---

Usage

import torch
import timm
from PIL import Image
import torchvision.transforms as T

# Load model
class DeepfakeDetector(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.backbone = timm.create_model("efficientnet_b4", pretrained=False, num_classes=0)
        self.head = torch.nn.Linear(self.backbone.num_features, 1)
    def forward(self, x):
        return self.head(self.backbone(x)).squeeze(-1)

model = DeepfakeDetector()
ckpt = torch.load("efficientnet_b4_deepguard_v2.pth", map_location="cpu")
model.load_state_dict(ckpt)
model.eval()

# Preprocess
transform = T.Compose([
    T.Resize((380, 380)),
    T.ToTensor(),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

img = Image.open("face.jpg").convert("RGB")
tensor = transform(img).unsqueeze(0)

with torch.no_grad():
    prob = torch.sigmoid(model(tensor)).item()

label = "FAKE" if prob > 0.5 else "REAL"
confidence = prob * 100 if label == "FAKE" else (1 - prob) * 100
print(f"{label} — {confidence:.1f}% confidence")

---

GradCAM Explainability

The model supports GradCAM heatmaps via hooks on model.backbone.blocks[-1] (the last MBConv block). Heatmaps highlight facial boundary regions, eye areas, and texture inconsistencies — the typical artifacts of face-swap deepfakes.

See the DeepGuard AI backend for the full GradCAM implementation.

---

Live Demo

Resource	Link
Web App	deep-guard-xai.vercel.app
API Backend	HuggingFace Space
Source Code	GitHub

---

Ethical Use

This model is developed for academic and research purposes only. It is intended to advance deepfake detection research, not to enable misuse. Non-consensual face swapping is strictly prohibited.

---

Citation

@misc{deepguard2026,
  title   = {DeepGuard AI: Explainable Deepfake Detection with EfficientNet-B4 and GradCAM},
  author  = {Arshad, Sowaiba},
  year    = {2026},
  url     = {https://huggingface.co/Sowaiba01/deepguard-ai}
}

---

License: MIT ·