☺️Image Protection Technology
Here are two main approaches:
Adding Invisible Watermarking and other copyright protection technologies to the works to facilitate rights protection.
While we cannot prevent AI from learning the works, we can increase the cost of AI learning the works to achieve a similar goal. Consider the following techniques: Adding Noise, Adversarial Perturbation, and Frequency Domain Masking.
Digital Watermarking: In our research, we found that merely resisting AI learning is not enough to protect your creativity, as there are always more robust models that can correct and learn from your images. We believe that digital watermarking is a crucial means to mark your work as your own and serve as a method for enforcing rights. The cost of removing watermarks is typically very high.
Adding Noise: By introducing random noise into the image, it becomes more challenging for AI to recognize the image's features. While this method is straightforward, excessive noise may be visually noticeable, and a limited amount of noise might not provide the desired disruption effect.
Adversarial Perturbation: This method specifically targets machine learning models to disrupt their behavior and increase their robustness. Adversarial perturbation involves adding specific small-scale noise to the image, causing the machine learning model to make incorrect predictions and preventing it from learning valuable information from the image. This method may have minimal visibility to the human eye, but high perturbation intensity can still degrade image details.
Frequency Domain Masking: By applying a mask on the image's frequency spectrum, the frequency information of the image is altered, confusing machine learning models.
Please note that implementing these protection techniques is not foolproof, but they can increase the difficulty for AI systems to exploit or modify creative works without proper authorization.
Last updated