My Art My Choice: Adversarial Protection Against Image Generation

Ilke Demir, Anthony Rhodes, Ram Bhagat, Nese Alyuz, Sinem Aslan, Umur A. Ciftci

Track: Short paper

Keywords: Image Generation, Adversarial Protection, Generative Models, Content Provenance

TL;DR: My Art My Choice helps artists to protect their content by learning an adversarial version which degrades style, content, and integrity of the generative model output.

Abstract: Generative AI enables realistic content production via publicly available interfaces, with diffusion models changing the creator economy through high-quality, low-cost content generation. However, artists are resisting unruly AI since their artwork is leveraged by such models without consent. My Art My Choice (MAMC) learns to generate adversarial "protected" versions of images that "break" diffusion models, with distortion levels controlled by the artist to balance fidelity vs. protection. We experiment across multiple datasets and image-to-image tasks, evaluating in visual, noise, structure, pixel, and generative spaces. A user study with 102 artists shows 98% willingness to use MAMC. Overall, MAMC offers crucial protection for preserving ownership against AI-generated content in a human-centric way.

Submission Number: 33