reality check

group pic

Reality Check is a robust AI-generated image detection prototype. It addresses the problem of identifying whether an image is authentic or AI-generated, even after common real-world reposting changes such as JPEG compression, blur, resizing, image noise, colour adjustment, and cropping. Users can upload an image and the system returns the likelihood of whether the image is AI-generated.

The project does not rely only on clean-image accuracy. It evaluates false positives, false negatives, robustness under transformations, and generalization to unseen data. This helps identify when the detector may incorrectly flag authentic images or miss realistic AI-generated images.

Development Tools

Visual Studio Code for development and project management
Git and GitHub for version control and collaboration
Python virtual environments for dependency management
Terminal/command-line tools for training, evaluation, packaging, and local deployment

Models and APIs Used

EfficientNet-B0 pretrained image-classification backbone
Binary classifier head with dropout and a sigmoid output for real-versus-AI prediction
FastAPI local inference API for image upload and prediction
Hugging Face Datasets / Hugging Face Hub tools for accessing and preparing SID Set data

Libraries and Frameworks Used

PyTorch and Torchvision for model building, training, and inference
timm for vision-model support
FastAPI and Uvicorn for the local web server
Pillow and OpenCV for image loading and preprocessing
Albumentations for robustness transformations and augmentation
NumPy and Pandas for data processing
Scikit-learn for evaluation metrics including ROC AUC, accuracy, precision, recall, and F1-score
Matplotlib for evaluation visualizations
PyYAML for experiment configuration

Datasets and Assets Used

SID Set: authentic, fully synthetic, and tampered image data. The project uses authentic and fully synthetic images for binary classification.
CIFAKE: CIFAR-10-based real and Stable Diffusion-generated images, used for training and evaluation.
WildFake: used as an external evaluation dataset to test generalization to unseen generators and image distributions.
Model checkpoints, prediction JSON files, evaluation reports, and error-analysis examples are generated locally and kept outside Git due to their size.