Stress-testing CLIP Image Detectors

A 120,000+ evaluation study measuring how CLIP-based AI-image detectors degrade under realistic image transformations.

Research question

AI-generated image detectors are often evaluated on clean benchmark data, while images in the real world are compressed, resized, cropped, blurred, noised, and captured through screenshots. I investigated how well a CLIP-based detector holds up after those transformations.

Experimental design

I designed and implemented a PyTorch and scikit-learn pipeline around CLIP ViT-B/32. The study covered ten transformations at five severity levels and completed more than 120,000 image evaluations.

Evaluation

Detector behavior was measured with AUROC, F1, precision, recall, and accuracy. Tracking each metric across increasing severity made it possible to distinguish gradual degradation from sharper failure points.

Outcome

The project quantified robustness gaps that clean-data evaluation can conceal. It was recognized among the Top 100 submissions in the National Student Research Institution Summer Research Hackathon 2026.

Tools

Python, PyTorch, scikit-learn, OpenAI CLIP, NumPy, Pandas, and reproducible experiment pipelines.

Related projects

Forma - Open-source Form Builder (Full-stack Project)

A production-grade form builder with conversational forms, AI generation, payments, automation, and real-time integrations.

Next.js, TypeScript, Amazon Bedrock, Stripe Connect, Docker, Resend

VECINTI Marketplace & Agent Platform

A production local-services marketplace combining a timezone-aware booking engine, geospatial discovery, agentic AI, and Stripe Connect infrastructure.

Next.js, TypeScript, PostgreSQL, PostGIS, Amazon Bedrock, Stripe Connect