Robust Detection of AI-Generated Images

Inspiration Seeing how easily standard AI detectors fail the second a synthetic image gets screenshotted, compressed, or shared on social media.

What it does A lightweight binary classifier that flags AI-generated images with 98.9% accuracy and a 0.9994 AUC, engineered to remain resilient against severe real-world degradations like JPEG compression and noise.

How we built it Fine-tuned a pretrained ConvNeXt-Tiny backbone via PyTorch on the CIFAKE dataset inside a Google Colab T4 environment, backed by an aggressive, physics-informed Albumentations pipeline.

Challenges we ran into Preventing the model from overfitting to pristine, high-resolution training artifacts and forcing it to learn robust structural cues that survive lossy re-encoding.

Accomplishments that we're proud of Achieving top-tier performance metrics (0.9994 AUC) while keeping the model strictly lightweight, and successfully executing batch inference across 20,000 test images.

What we learned Data augmentation pipelines tailored to real-world physical distortions are far more critical for generalization than raw model capacity or training from scratch.

What's next for Robust Detection of AI-Generated Images Expanding the augmentation suite to handle multi-format video deepfakes and deploying the pipeline as a real-time browser extension for instant verification.