🛡️ RealityGuardAI: Complete Deepfake Detection Architecture

RealityGuardAI is an ultra-fast, highly optimized AI engine specifically designed to detect visual artifacts, behavioral anomalies, and audio-visual desynchronization commonly found in AI-generated Deepfake videos and images.

This Model Repository specifically hosts the Lightweight Artifact CNN weights, which act as the core deep learning visual detection module of the larger RealityGuardAI pipeline.

🏗️ Full Pipeline Architecture

This CNN operates at Stage 7.3 of our massive multimodal pipeline. To view, run, and experiment with the entire pipeline (including Lip-Sync Detection, Blink Analysis, Handcrafted Feature Extraction, and FastAPI integrations), please visit the official GitHub repository:

👉 Satyam-123336/DeepFake-Detection on GitHub

---

🧠 Model Details (Lightweight Artifact CNN)

• Architecture: 3-Layer Sequential CNN with Adaptive Average Pooling and a 2-Layer MLP Head.
• Parameters: ~100KB (Highly optimized for edge-device CPU inference).
• Task: Binary Classification (0 = Real, 1 = Fake).
• Format: PyTorch State Dictionary (.pt).

🎯 Intended Use & Capabilities

This Model Repository hosts the weights specifically for the deep learning component (Stage 7.3) of RealityGuardAI.

While the full pipeline uses classical OpenCV algorithms to detect lighting asymmetry and texture spread (Stage 7.2), this PyTorch CNN model is specifically designed to analyze raw facial crops to detect:

• Generative AI artifacts (GAN/Diffusion traces and upscaling noise)
• Deep pixel-level blending inconsistencies (where a fake face is stitched onto a real head)
• Structural anomalies generated by deepfake rendering engines

💻 How to Use

You can easily integrate this CNN into your own PyTorch projects using the huggingface_hub library. It will automatically download the weights if they are missing locally.

import torch
from huggingface_hub import hf_hub_download
from torch import nn

# 1. Define the Lightweight Architecture
class LightweightArtifactCNN(nn.Module):
    def __init__(self, num_classes: int = 2):
        super().__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(2),
            nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(2),
            nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1),
            nn.ReLU(inplace=True),
            nn.AdaptiveAvgPool2d((1, 1)),
        )
        self.classifier = nn.Sequential(
            nn.Flatten(),
            nn.Linear(64, 32),
            nn.ReLU(inplace=True),
            nn.Dropout(0.2),
            nn.Linear(32, num_classes),
        )
    def forward(self, x):
        return self.classifier(self.features(x))

# 2. Auto-Download and Load Weights
model = LightweightArtifactCNN()
weight_path = hf_hub_download(repo_id="Satysam-26/RealityGuardAI", filename="lightweight_artifact_cnn_1000_fresh.pt")
model.load_state_dict(torch.load(weight_path, map_location="cpu"))
model.eval()

print("Model successfully loaded and ready for inference!")

🌱 Environmental Impact

Unlike massive billion-parameter transformer models, the RealityGuardAI CNN is purposefully designed to be executed entirely on standard CPUs.

• Hardware Type: CPU-Optimized
• Carbon Emitted: Negligible compared to standard GPU-bound Video AI systems.

⚠️ Limitations & Bias

• Resolution Sensitivity: The CNN is optimized for well-lit, clearly visible facial crops. Heavy compression artifacts (like 144p WhatsApp videos) may increase false-positive rates.
• Audio-Visual Dependency: For full robustness, this visual model must be fused with our behavioral and lip-sync models found in the main GitHub repository.

📚 Citation

If you use this architecture or model in your research, please cite it using the following DOI:

@software{realityguardai_2026,
  author       = {Satyam},
  title        = {RealityGuardAI: A Multi-Signal and Explainable Deepfake Detection Engine},
  year         = 2026,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.22282866},
  url          = {https://doi.org/10.5281/zenodo.22282866}
}