Making AI-Generated Content Traceable with a Hidden Code

Intel's patent describes a method for embedding a hidden, verifiable authentication code directly into AI-generated content, like deepfakes, during their creation process.

What it covers

The patent describes a system that uses three interconnected AI networks: a generative neural network (the "creator"), a discriminator neural network (the "critic"), and an authenticator neural network (the "verifier"). The creator makes new samples, trying to fool the critic into thinking they are real. Crucially, the verifier also "digests" both real and created content, along with a secret "authentication code." The verifier then helps embed this code into the created samples by influencing the creator's training process (specifically, its "generator loss"). This results in AI-generated content that contains a hidden, verifiable authentication code, which the verifier can later use to confirm the content's origin. For example, a video generated by this system would have an invisible digital watermark proving it came from a specific source, verifiable by the authenticator neural network based on the embedded code.

What it doesn't cover

• —Does not cover methods for detecting deepfakes after they have been generated without an embedded authentication code.
• —Does not cover authentication systems that rely on external watermarking or metadata added after content generation.
• —Does not cover AI-generated content that does not involve a generative adversarial network (GAN) architecture.
• —Does not cover authentication methods that do not involve an "authenticator neural network" directly contributing to the generator's training loss.

The clever bit

The novelty lies in integrating the authenticator directly into the GAN's training loop. Instead of just trying to fool a discriminator, the generative network is also trained to embed a specific, verifiable authentication code, making the content inherently traceable from its origin.

Why it matters

This technology addresses the growing challenge of distinguishing between real and AI-generated content, especially deepfakes. By embedding an authentication code during creation, it offers a way to verify the source of digital media. This could be crucial for combating misinformation and ensuring trust in digital content, particularly in news, social media, and legal contexts. It provides a mechanism for content creators to prove the authenticity and origin of their AI-generated works.

Real-world examples

• 1.Verifiable AI-generated news reports
• 2.Authentic AI-created product advertisements
• 3.Secure synthetic data for training other AI models
• 4.Digital watermarking for AI-generated art