Inspiration
Generative AI poses real risks to digital trust, but standard AI detectors fail the moment an image is compressed, cropped, or blurred on social media. We wanted to build a detector optimized for real-world robustness, not just pristine lab conditions.
What it does
It evaluates whether an image is AI-generated or authentic, remaining accurate even after aggressive post-processing. Instead of a black-box answer, it provides transparent, rule-based explanations in JSON format and sorts images into actionable moderation tiers via an interactive HTML dashboard.
How we built it
We fused a CLIP ViT-B/32 semantic branch with a custom 2D Fast Fourier Transform (FFT) CNN frequency branch. This dual-branch architecture was trained with aggressive data augmentation and paired with a custom JPEG balance checker to prevent shortcut learning.
Challenges we ran into
Our biggest hurdle was preventing the model from simply learning "JPEG = Real." We had to implement a balance checker and apply heavy compression augmentations to break this bias. We also had to carefully tune the branches so the model wouldn't collapse when heavy blur wiped out high-frequency signals.
Accomplishments that we're proud of
We achieved competitive real-world robustness with a model strictly under 2 billion parameters. We also successfully connected a PyTorch backend to a clean, live HTML frontend that delivers actionable policy recommendations instead of just a raw probability score.
What we learned
We learned that upsampling low-resolution datasets (like CIFAKE) invents synthetic frequencies that pollute the FFT spectrum. We also realized that analyzing compression quantization tables is essential before trusting any AI detection benchmark.
What's next for Team UM - AI-Image-Detector
We plan to expand to 3-channel $YCbCr$ frequency processing, integrate LoRA for deeper fine-tuning, and upgrade our dashboard with a Python API to support real-time, drag-and-drop image uploads.
Log in or sign up for Devpost to join the conversation.