🔍 Detect deepfakes with state-of-the-art accuracy
A ResNeXt-101 32×8d backbone — initialised from Instagram's weakly-supervised pretrained weights — fine-tuned to expose AI-generated and manipulated faces with high confidence.
✅ Real Face 🚨 Deepfake
────────────── ──────────────
Confidence: 97.3% Confidence: 99.1%
🗺️ Navigation
📊 Overview · 🏗️ Architecture · 🚀 Quick Start · 🖼️ Examples · 🏋️ Training · ⚠️ Limitations · 📜 Cite
📊 Model Overview
|
|
🏗️ Architecture
The backbone uses grouped convolutions with cardinality 32 — each layer splits into 32 parallel transformation paths, then aggregates. This lets the network learn diverse artefact patterns (blending seams, frequency inconsistencies, unnatural textures) simultaneously.
┌─────────────────────────────────────────────────────────────┐
│ ResNeXt-101 32×8d │
├─────────────────────────────────────────────────────────────┤
│ │
│ 📷 Input ──▶ 🌱 STEM │
│ Conv 7×7 │ BN │ ReLU │ MaxPool │
│ 3 → 64 channels │
│ │ │
│ ┌────────▼────────┐ │
│ │ 🧩 LAYER 1 │ ×3 blocks · ch 256 │
│ └────────┬────────┘ │
│ ┌────────▼────────┐ │
│ │ 🧩 LAYER 2 │ ×4 blocks · ch 512 │
│ └────────┬────────┘ │
│ ┌────────▼────────┐ │
│ │ 🧩 LAYER 3 │ ×23 blocks · ch 1024 ◀── deepest │
│ └────────┬────────┘ │
│ ┌────────▼────────┐ │
│ │ 🧩 LAYER 4 │ ×3 blocks · ch 2048 │
│ └────────┬────────┘ │
│ Global Avg Pool │
│ │ │
│ ┌─────────▼─────────┐ │
│ │ 🎯 FC HEAD │ 2048 → num_classes │
│ └───────────────────┘ │
└─────────────────────────────────────────────────────────────┘
Each bottleneck block:
┌──────────────────────────────────────────────────┐
│ 1×1 Conv (expand) → 3×3 GroupConv (groups=32) │
│ → 1×1 Conv (compress) + Skip Connection │
└──────────────────────────────────────────────────┘
🚀 Quick Start
1️⃣ Install dependencies
# Clone the repo
git clone https://github.com/accel-reg/deepfake-detection.git
cd deepfake-detection
# Install requirements
pip install -r requirements.txt
📦 What's in requirements.txt?
torch>=1.13
torchvision>=0.14
Pillow
opencv-python
huggingface_hub
2️⃣ Load the model
import torch
from model import DeepfakeDetector # from the repo
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# Option A — load local file
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
model = DeepfakeDetector()
model.load_state_dict(torch.load("ig.bin", map_location="cpu"))
model.eval()
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# Option B — pull from HuggingFace 🤗
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="accel69/depfake-detection", filename="ig.bin")
model = DeepfakeDetector()
model.load_state_dict(torch.load(path, map_location="cpu"))
model.eval()
3️⃣ Preprocess & predict
from torchvision import transforms
from PIL import Image
# ── Standard ImageNet preprocessing ───────────────────────────
transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406],
std =[0.229, 0.224, 0.225]
),
])
# ── Predict ────────────────────────────────────────────────────
img = Image.open("face.jpg").convert("RGB")
x = transform(img).unsqueeze(0) # → (1, 3, 224, 224)
with torch.no_grad():
probs = torch.softmax(model(x), dim=1)
pred = probs.argmax(dim=1).item()
label = "🚨 FAKE" if pred == 1 else "✅ REAL"
confidence = probs[0, pred].item()
print(f" Result : {label}")
print(f" Confidence : {confidence:.2%}")
🖼️ Inference Examples
📁 Batch inference on a folder
from pathlib import Path
image_dir = Path("frames/")
results = {"real": 0, "fake": 0}
for img_path in sorted(image_dir.glob("*.jpg")):
img = Image.open(img_path).convert("RGB")
x = transform(img).unsqueeze(0)
with torch.no_grad():
probs = torch.softmax(model(x), dim=1)
is_fake = probs.argmax().item() == 1
confidence = probs.max().item()
label = "🚨 FAKE" if is_fake else "✅ REAL"
results["fake" if is_fake else "real"] += 1
print(f" {img_path.name:<35} {label} ({confidence:.2%})")
print(f"\n 📊 Summary — ✅ Real: {results['real']} | 🚨 Fake: {results['fake']}")
⚡ GPU acceleration
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
print(f" 🔥 Running on : {device}")
print(f" ⚡ CUDA cores : {torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'N/A'}")
# Move input to same device
x = x.to(device)
with torch.no_grad():
probs = torch.softmax(model(x), dim=1)
🎥 Video frame-by-frame analysis
import cv2
cap = cv2.VideoCapture("video.mp4")
fake_frames = 0
total_frames = 0
print(" 🎬 Analysing video...")
while cap.isOpened():
ret, frame = cap.read()
if not ret:
break
img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
x = transform(img).unsqueeze(0)
with torch.no_grad():
pred = model(x).argmax(dim=1).item()
fake_frames += pred
total_frames += 1
cap.release()
fake_pct = fake_frames / total_frames
verdict = "🚨 LIKELY DEEPFAKE" if fake_pct > 0.5 else "✅ LIKELY REAL"
print(f"\n ┌─────────────────────────────────┐")
print(f" │ 🎬 Total frames : {total_frames:<12}│")
print(f" │ 🚨 Fake frames : {fake_frames:<12}│")
print(f" │ ✅ Real frames : {total_frames-fake_frames:<12}│")
print(f" │ 📊 Fake ratio : {fake_pct:<11.1%} │")
print(f" │ 🏁 Verdict : {verdict:<12}│")
print(f" └─────────────────────────────────┘")
🏋️ Training Details
Training Pipeline
══════════════════════════════════════════════════════
📦 Backbone Instagram WSL ResNeXt-101 32×8d
🖼️ Resolution 224 × 224 RGB
📐 Normalisation ImageNet mean [0.485 0.456 0.406]
std [0.229 0.224 0.225]
📉 Loss function Cross-Entropy
🔄 Augmentation Horizontal flip · Colour jitter
Random crop · Rotation
| ⚙️ Hyperparameter | 📋 Value |
|---|---|
| 🧱 Backbone init | Instagram WSL pretrained (WSL-Images) |
| 📷 Input resolution | 224 × 224 |
| 📐 Normalisation | ImageNet mean / std |
| 📉 Loss | Cross-Entropy |
| 🔄 Augmentations | Flip, colour jitter, random crop |
📖 Full configs, dataset prep scripts and training logs → GitHub Repository
⚠️ Limitations
🚧 Read before deploying in any production or real-world system.
| ⚠️ Risk | 📋 Details |
|---|---|
| 🆕 Novel forgery methods | May not detect unseen GAN/diffusion techniques |
| 📐 Alignment sensitivity | Poor face crop → lower accuracy. Use a dedicated face detector first |
| 🌍 Distribution shift | Different cameras, compression, or lighting may degrade results |
| ⚖️ Demographic bias | Not audited across demographic groups — evaluate independently |
| 🔁 No temporal context | Frame-level only — no multi-frame consistency modelling |
🎯 Intended Use
📜 License & Citation
Released under the MIT License — free to use, modify, and distribute with attribution.
@misc{ig-deepfake-detection-2025,
author = {accel69},
title = {ig.bin — Deepfake Face Detection with ResNeXt-101 32x8d},
year = {2025},
publisher = {HuggingFace},
url = {https://huggingface.co/accel69/depfake-detection}
}
