#NeurIPS2024 is looking like the AI Hogwarts. Or is it just me?

Presenting the all-woman team behind Semi-Truths, (+ @polochau.bsky.social ) landed at #NeurIPS2024!!

🤗 Come chat with us tomorrow at 11am - 2pm, West Ballroom A-D #5211.

Learn about the weaknesses of AI-Generated Image Detectors!

Every image is enriched with attributes quantifying the magnitude of change achieved. Evaluating performance on these attributes provides insights into detector biases.

💡 UniversalFakeDetector suffers >35 point performance drop on different scenes, and >5 points on magnitude of change.

🔧 To control what is changed in an image and how, we use semantic segmentation datasets that provide real images, entity masks, and entity labels.

We perturb entity & image captions with LLMs, then apply different diffusion models and augmentation techniques to alter images.

🚀 We present Semi-Truths, a dataset for the targeted evaluation and training of AI-Augmented Image Detectors.

It includes a wide array of scenes & subjects, as well as various magnitudes of image augmentation. We define “magnitude” by size of the augmented region and the semantic change achieved.

An attacker may keep most of the original image, and only change a localized region to drastically change the narrative!

🔍 One such case is known as “Sleepy Joe”, where a video of Joe Biden was changed only in the facial region to make it appear as though he fell asleep at a podium.

🚨 NeurIPS 2024 🚨How robust are AI-Generated Image Detectors?

🤔 Can they detect various magnitudes of image augmentations?
💡 Does performance fluctuate across scenes?

🚀 Find out with Semi-Truths: 1.5 million images for the targeted evaluation of AI-generated images. arxiv.org/abs/2411.07472