"Steg.AI's unique watermark technology puts tech giants to shame"

2023-08-04

While companies like Meta, Microsoft, Stability AI, and Midjourney are working hard, Steg.AI is leading the way.

In May 2023, the world witnessed the shocking image of the Pentagon engulfed in smoke. Many news channels reported on this event based on these images, and even the stock market reacted by plummeting in a short period of time. Later, it was proven that this was a fake image generated by artificial intelligence.

These events highlight the challenges of identifying AI-generated content in various situations. They have also sparked a larger discussion about deepfakes and fake images, thanks to the capabilities of generative AI tools to generate realistic images based on prompts.

All of this has led governments around the world to seek solutions and regulatory methods to secure AI. Prominent AI companies, including Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI, have voluntarily committed to taking measures such as adding watermarks to AI-generated content to enhance its security, but there has been no substantial progress so far.

Current methods, such as encoding data into images or audio, can be easily bypassed. Therefore, a powerful and invisible watermark that is easy to apply and detect, while resistant to transformations, is necessary, as research has shown that it is difficult for people to distinguish between human-generated and AI-generated content. The ability to prove the source of content creation has become increasingly important due to rampant online intellectual property theft.

Meanwhile, California-based platform Steg.AI has developed a deep learning-based solution that seamlessly embeds imperceptible watermarks into digital content. Even if the image is modified, compressed, or manipulated, Steg.AI's watermark remains intact. These watermarks are very robust and can even be captured using an iPhone camera on a screen or when printed.

Steg.AI's watermark solution has applications in various scenarios, such as stock photography services, content sharing on platforms like Instagram, pre-release copies of movies, and protecting confidential documents. The development of their early products faced challenges, which led them to focus on robustness, a prominent feature that is favored by customers.

How It Works

The core concept of Steg.AI is to seamlessly integrate watermarks into AI-generated images before they are published. Although the specific process is confidential, the basic idea revolves around a pair of machine learning models. One model customizes the placement of the watermark in the image, ensuring it is imperceptible to the human eye while detectable by decoding algorithms.

Similar to an almost invisible and minimally alterable QR code, this approach potentially accommodates several kilobytes of data, enough for URLs, hash functions, and plain text messages. Each page of a multi-page document or video frame can hide different codes, multiplying the capacity.

The company's extensive work can be traced back to a CVPR paper in 2019 and acquisitions of Phase I and Phase II SBIR government funding. Co-founders Eric Wengrowski and Kristin Dana have been involved in academic research before, dedicating years to refining their methods.

Steg.AI's progress has been supported by NSF grants and a total of $1.2 million in angel investments. Recently, the company announced a significant milestone with a $5 million seed funding round led by Paladin Capital Group, with participation from Washington Square Angels, NYU Innovation Venture Fund, and individual angel investors.

What Major Tech Companies Are Doing

Major tech companies have taken steps to apply watermark technology to their content. Microsoft announced at its annual Build conference the addition of new media attribution capabilities in Bing Image Creator and Designer, allowing users to verify AI-generated images and videos. This innovation involves encryption methods to tag and sign content with metadata indicating its source. Websites must adopt the Content Authenticity and Provenance Alliance (C2PA) specifications developed jointly with Adobe, Arm, Intel, Microsoft, and Truepic.

However, the impact of Microsoft's efforts depends on the adoption of widespread media origin standards and has received support from companies like Stability AI and Google, which are also exploring similar approaches. Shutterstock and Midjourney have already adopted guidelines to embed markers indicating the origin of AI-generated content.

On the other hand, collaborative research by Meta AI, the International Research Center at Rensselaer Polytechnic Institute, and Sorbonne University has developed an innovative technology that seamlessly integrates watermarks into the image generation process while preserving the integrity of the architecture. This method modifies pre-trained generative models to effectively integrate watermarks into the generated images, enhancing security and computational efficiency. This technology enables model providers to distribute model versions with different watermarks to different user groups, facilitating ethical usage monitoring.

This technology is highly valuable for media organizations in identifying computer-generated images. Researchers successfully combined watermarks with generative models with minimal adjustments using the Latent Diffusion Model (LDM). This process involves fine-tuning the LDM decoder using perceptual image loss and a simplified deep watermarking method called HiDDeN. The technology demonstrates strong performance in image editing tasks, even in severely cropped images, while preserving the utility of the original model in various LDM-based tasks.

Last month, the actions of these seven major tech companies supported the Biden administration's efforts to strengthen regulation of the booming and popular AI technology. The US Congress is also considering a bill that requires disclosure of AI involvement in political advertisements.

RECENT AI NEWS

RECENT AI TOOLS