Image Anti-Spoofing System
Detects whether an image is real or fake (AI-generated OR deepfake/manipulated).
Third project in a multi-modal anti-spoofing portfolio, alongside Voice-Anti-Spoofing and Text-Anti-Spoofing.
Architectures
Three architectures were trained and compared on the same data:
| Architecture | Description |
|---|---|
BaselineCNN | Shallow 3-block Conv2D CNN |
DeeperCNN | Deeper 4-block CNN with BatchNorm |
ViTModel | ViT-B/16 (ImageNet-pretrained) adapted for RGB images via transfer learning |
Training data
prithivMLmods/AI-vs-Deepfake-vs-Real — original 3-way labels (Artificial, Deepfake, Real) merged into a binary task: Real (0) vs Fake (1, covering both AI-generated and deepfake images), to detect AI manipulation of any kind in one model.
Results (test set accuracy)
| Model | Test Accuracy | File size |
|---|---|---|
| Baseline CNN | 98.80% | ~26MB |
| Deeper CNN | 99.20% | ~26MB |
| ViT | 100.00% | ~343MB |
Important caveat on the ViT's 100% score: a perfect test-set score is a signal to scrutinize, not just celebrate. It likely reflects that this dataset's fake images (from a specific, fixed set of generators) have consistent, learnable artifacts, rather than the model having solved "detect any AI-manipulated image" universally. Performance against newer/unseen generators (e.g. Midjourney v6, Flux, Stable Diffusion 3, or generators not represented in this training set) is untested and likely to be lower - this is a common "concept drift" issue in this field, since fake-image generators keep improving and older detectors don't automatically generalize to them.
Recommended checkpoint for deployment: deeper_cnn_model.pth. At 99.20% accuracy and ~26MB (vs. ViT's 343MB), it offers a strong practical tradeoff between accuracy and deployment cost (faster cold starts, less bandwidth) with a more plausible, less potentially-overfit accuracy figure than the ViT's perfect score.
Files in this repo
baseline_cnn_model.pthdeeper_cnn_model.pth(recommended)vit_model.pthmodel.py- architecture class definitions + preprocessing pipeline, required to load any of the above
Input format
- Any image
PILcan open (.jpg,.png, etc.) - Resized to 224x224
- Normalized with standard ImageNet mean/std
model.py includes a preprocess_image() function that handles all of this automatically.
Usage
from huggingface_hub import hf_hub_download
import sys
model_py_path = hf_hub_download(repo_id="LaabhGupta/image-antispoofing", filename="model.py")
weights_path = hf_hub_download(repo_id="LaabhGupta/image-antispoofing", filename="deeper_cnn_model.pth")
sys.path.insert(0, model_py_path.rsplit("/", 1)[0])
from model import load_model, predict
model = load_model("deeper", weights_path, device="cpu")
label, confidence = predict("path/to/image.jpg", model, device="cpu")
print(label, confidence)