Foxglove Protects Artwork from AI Training with On-Device Image Poisoning

Summary

Foxglove is an image-protection app for artists that processes images on the user’s device before they are shared. It removes EXIF, GPS, camera, and software metadata; can embed a hidden ownership message in image pixels; and adds adversarial perturbations intended to corrupt how AI models learn from scraped training data. The product emphasizes that poisoning affects training rather than inference, so a poisoned image may still be described normally by ChatGPT or another model. The hidden message uses block-level steganographic encoding and is best preserved in PNG, although aggressive social-media recompression may remove it. Foxglove says the perturbation is more persistent through compression, while its hidden-message system avoids flat regions when there is not enough texture to conceal the data. Users can apply the protections to batches of up to 20 images, add a visible watermark, limit resolution, and include an artist identity in the hidden message. The app is free, available on iPhone, and includes a verifier for revealing embedded messages. Planned work includes style-cloaking protection, an Android version, JPEG-resilient steganography, and model-specific perturbations; these features are described as in progress or under research.