AI image watermarking has recently faced a major setback. A new tool called UnMarker claims to be able to crack almost all AI image watermarks in just 5 minutes, including Google's highly praised HiDDeN watermarking technology, which has been completely broken, and even the SynthID watermarking system, considered more secure, is facing a high attack success rate of 79%.
This technological breakthrough has caused a stir in the tech world, prompting the industry to re-examine the security and reliability of existing watermarking technologies. Traditional watermarks have long been seen as an important barrier for protecting creators' intellectual property, especially with the increasing popularity of AI image generation technology, making their importance even more prominent.
Differing from common visible watermarks, the hidden watermark technology used in AI images embeds identification information within the deep data structure of the image, particularly in the spectral features. These spectral features consist of two parts: spectral magnitude and phase. The watermark embedding process mainly achieves this by modifying the spectral magnitude, which allows the watermark to remain relatively stable against common operations such as image cropping, blurring, and compression.
André Kassis, the developer of UnMarker, expressed strong confidence in this technology. He stated that the core advantage of UnMarker lies in not requiring precise positioning of the watermark, but rather directly modifying the spectral information of the image. By effectively disrupting the watermark signal, it can achieve the removal effect. This general design enables UnMarker to adapt to various types of watermarking systems, and experimental data shows that its removal success rate fluctuates between 57% and 100%.
However, UnMarker also has certain limitations. During the watermark removal process, the tool may cause minor modifications to the original image, but Kassis pointed out that by appropriately cropping the image, a better processing result can be achieved. More importantly, UnMarker can run normally on consumer-grade graphics cards, greatly reducing the usage threshold for ordinary users and making watermark removal technology more accessible.
Microsoft's latest research data shows that the accuracy of ordinary users in identifying AI-generated images is only 62%, further highlighting the important role of watermarking technology in AI content identification. However, the emergence of tools like UnMarker is posing a serious challenge to the originally considered secure and reliable watermark protection system.
This technological trend reveals an ongoing "battle of offense and defense" in the field of AI content protection. As de-watermarking technology continues to advance, how to establish a more robust content identification and copyright protection mechanism has become a key issue that the entire AI industry needs to address urgently.
Project address: https://github.com/andrekassis/ai-watermark
