EfficientNet-B4 Deepfake Detector with Grad-CAM Explainability

A high-accuracy deepfake face detector trained on Celeb-DF v2, combining an EfficientNet-B4 backbone with Grad-CAM spatial attribution and a deterministic forensic report generator. The model classifies face images as real or fake and highlights which facial region triggered the decision.

Bachelor project — Sapienza Università di Roma, AI & Applied Computer Science

Model Performance

Video-level scores are computed by mean-aggregating frame probabilities per video ID, which suppresses single-frame noise and reflects real-world deployment.

What Makes This Different

• Explainable predictions — Grad-CAM heatmaps highlight the exact facial zone (forehead, eyes, nose, jaw, or hairline) that triggered the detection.
• Forensic text output — A template engine converts confidence + activated zones into a structured human-readable forensic report (4 confidence tiers).
• Video-level reasoning — Frame scores are aggregated per video for a single robust verdict.
• Interactive demo — Gradio app supports both image and video input.

Architecture

Input (224×224 face crop)
  └─ EfficientNet-B4 backbone (ImageNet pretrained)
       ├─ Blocks 0–4  →  frozen (feature extraction)
       └─ Blocks 5–8  →  fine-tuned (LR = 1e-4)
           └─ Global Average Pooling
               └─ Dropout(0.4) → Linear(1792→256) → ReLU → Dropout(0.2) → Linear(256→1)
                   └─ Sigmoid → probability [0, 1]  (≥ 0.5 = Fake)

• Loss: Focal Loss (α=0.25, γ=2.0) — handles the 5:1 fake/real imbalance
• Optimizer: AdamW with differential learning rates (backbone 1e-4, head 5e-4)
• Scheduler: CosineAnnealingLR over 20 epochs with early stopping (patience=5)
• GPU: NVIDIA RTX A4000

Dataset

Celeb-DF v2 — 590 real celebrity videos + 5,639 high-quality deepfake videos.

• 15 frames extracted per video (uniform temporal sampling)
• MTCNN face detection → 224×224 crops, 20 px margin
• Split by video ID (80/10/10) — prevents identity leakage between train and test
• ~74,000 real face crops · ~477,000 fake face crops

Usage

Quick inference (image)

import torch
from torchvision import transforms
from PIL import Image
from huggingface_hub import hf_hub_download

# Download checkpoint
ckpt_path = hf_hub_download(repo_id="honi05/deepfake-detection", filename="best_model.pt")

# Load model
from src.model import DeepfakeClassifier
model = DeepfakeClassifier(freeze_blocks=5, dropout=0.4, backbone='b4')
state = torch.load(ckpt_path, map_location="cpu", weights_only=True)
model.load_state_dict(state["model_state_dict"])
model.eval()

# Preprocess
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])

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

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

print(f"Fake probability: {prob:.3f}")
print("Verdict:", "FAKE" if prob >= 0.5 else "REAL")

Grad-CAM explainability

from src.gradcam import GradCAM

grad_cam = GradCAM(model)
heatmap, confidence = grad_cam.compute(img_tensor)   # (224,224) heatmap in [0,1]
overlay = grad_cam.overlay(img_pil, heatmap)         # PIL image with jet overlay

top_zones = grad_cam.get_top_zones(heatmap, top_k=2)
print("Most activated zones:", top_zones)

Forensic report

from src.forensic_text import generate_forensic_report

report = generate_forensic_report(confidence=0.91, zone1="eyes", zone2="jaw")
print(report)
# HIGH CONFIDENCE FAKE (91.0%) — Eyes region shows unnatural reflection/texture
# patterns inconsistent with genuine facial geometry. Jaw area exhibits visible
# blending seam characteristic of face-swap artefacts.

Gradio demo (image + video)

python demo/app.py

Explainability — Facial Zones

The model maps Grad-CAM activations to 5 facial zones (pixel rows in the 224×224 crop):

The top-2 activated zones are included in the forensic report.

Ablation Results

Key findings: data augmentation and the larger B4 backbone provide the biggest gains. Focal loss measurably improves handling of the class imbalance. Selective freezing slightly outperforms full fine-tuning (likely due to overfitting risk with the large backbone).

Limitations

• Binary classification only (real vs. fake) — does not identify the generation method
• No temporal modelling — each frame is classified independently
• Trained on Celeb-DF v2 only — may not generalise equally to StyleGAN or diffusion-based fakes
• High-compression video can suppress the artefacts the model relies on
• False Positive Rate of ~15.9% on the test set

Files

Full source code: github.com/Honi05/DeepFakeDetector

Citation

If you use this model, please cite:

@misc{arora2026deepfake,
  title   = {Deepfake Detection with Explainable Forensic Analysis Using EfficientNet-B4 and Grad-CAM},
  author  = {Arora, Honi},
  year    = {2026},
  url     = {https://huggingface.co/honi05/deepfake-detection}
}

Acknowledgements

• Celeb-DF v2 — Li et al., CVPR 2020
• EfficientNet — Tan & Le, ICML 2019
• Grad-CAM — Selvaraju et al., ICCV 2017
• Focal Loss — Lin et al., ICCV 2017
• facenet-pytorch — MTCNN implementation