PhotoGuard: The Ultimate Protection Measure Against Unauthorized Image Processing

2023-07-31

MIT has proposed a new artificial intelligence tool aimed at preventing unauthorized modifications to images, such as those made by models like DALL-E and Midjourney. From romantic poetry to Salvador Dali-style images, generative AI can now do it all. Its performance is often indistinguishable from human creations. Since the Turing test established imitation of humans as the standard, the discussion about technology imitating humans has been a major topic of public debate. In the risk of potential misuse of the technology, the community has been trying to distinguish between human-created text and text generated by artificial intelligence. MIT's solution: PhotoGuard Scientists from MIT's Computer Science and Artificial Intelligence Laboratory have developed a new AI tool called "PhotoGuard" that aims to prevent unauthorized modifications to images. PhotoGuard utilizes "adversarial perturbations," a method of making small changes to pixel values that are imperceptible to the human eye but can be detected by computer models. These perturbations disrupt the AI model's ability to process images effectively. PhotoGuard uses two attack methods to generate these perturbations. The "encoder" attack targets the AI model's potential representation of an image, causing the model to perceive the image as random. The goal of this attack is to disrupt the Latent Disentangled Model (LDM), which encodes the input image into a latent vector representation that is then used in the process of generating new images. They achieve this goal by solving an optimization problem using projected gradient descent (PGD). The perturbation added to the original image results in the LDM generating an irrelevant or unrealistic image. On the other hand, the "diffusion" attack defines a target image and optimizes the perturbation to make the final image closely resemble the target image. This attack is more complex and aims to disrupt the diffusion process itself, targeting not only the encoder but also the complete diffusion process with text prompt conditioning. Its goal is to generate a specific target image (e.g., random noise or grayscale image) by solving another optimization problem using PGD. This attack not only nullifies the immunity of the image but also nullifies the effect of the text prompt. Hadi Salman, the lead author of this research paper and a PhD student at MIT, told AIM, "Essentially, PhotoGuard adds an additional layer of protection to images against tampering by diffusion models." By leveraging these imperceptible pixel modifications to the human eye, PhotoGuard protects images from manipulation by such models. For example, consider an image containing multiple faces. You can mask the faces you don't want to be modified and use the prompt "two people attending a wedding." Upon submission, the system will adjust the image accordingly, creating a plausible depiction of two people attending a wedding ceremony. Now, consider protecting the image from editing; adding perturbations to the image before uploading can prevent it from being modified. In this case, the final output lacks realism compared to the original non-immune image. Salman concludes, "I am skeptical about AI replacing human creativity. I expect AI to become another (powerful) tool in the hands of designers to enhance individual productivity and better express their ideas in the face of technical barriers." Decoding the problem Recent discussions in the Senate regarding AI regulation have focused on the most pressing issues, such as copyright and artist incentives. CEOs of companies like OpenAI, HuggingFace, and Meta have testified before the US Congress, offering suggestions on the potential dangers of AI and proposing the establishment of a new government agency to issue licenses for large AI models, revoke non-compliant licenses, and set safety protocols. The main driving force behind this call for regulation stems from concerns about copyright infringement. The artist community first filed lawsuits against image generation companies like Stability AI, Midjourney, and DeviantArt, seeking compensation for the unauthorized use of their artwork. AI-generated content has faced opposition from image companies like Shutterstock and Getty, as well as artists who view it as a threat to their intellectual property. However, in the end, most of them have established partnerships with the relevant companies. Adobe's Firefly is a generative image generator designed for "safe commercial use." Adobe provides intellectual property compensation protection to ensure users do not face legal issues when using it. It is based on Nvidia's Picasso, which was trained on licensed images from Getty Images and Shutterstock. Shutterstock also collaborates with OpenAI, the creator of DALL-E, to provide training data. It also offers comprehensive compensation protection for enterprise customers using generative AI images on its platform to ensure protection against any potential legal claims related to image usage. Google, Microsoft, and OpenAI have also started mitigating copyright issues by adding watermarks. Source: https://analyticsindiamag.com/meet-photoguard-the-ultimate-shield-against-unauthorised-image-manipulation/

RECENT AI NEWS

RECENT AI TOOLS