Blog

The rise of deepfake technology has brought a new set of challenges to the digital world. In 2023 alone, the total number of deepfake videos online reached 95,820, marking a staggering 550% increase since 2019. Of these, a worrying 98% constitute deepfake pornography, overwhelmingly targeting women, with 99% of individuals featured in these videos being female​​​​. 

This explosion of deepfakes not only challenges the concept of digital consent but also raises critical concerns about misinformation and the misuse of AI technology. As these AI-generated creations become more sophisticated, distinguishing between what's real and what's artificial is becoming increasingly difficult. This blurring of lines poses significant risks, from spreading false information to infringing on individual rights.

In an attempt to overcome these challenges, Google DeepMind introduced SynthID. It’s a tool aimed at watermarking and identifying AI-generated content, making it easier to identify their AI origins. SynthID embeds a digital watermark directly into the pixels of AI-generated images or audio, creating a marker that is imperceptible to humans but detectable for identification​​. This innovation addresses a key challenge: the brittleness of traditional watermarking methods, which often fail when images are cropped, resized, or edited. 

By offering a way to verify the authenticity of digital content, SynthID emerges as a vital tool in maintaining the integrity and trustworthiness of online media. 

How SynthID Works 

The tool operates using two specialized deep-learning models: one to watermark AI-generated content and the other to identify it. 

Watermarking: The first model alters the pixels in an image to insert an invisible watermark. This watermark doesn't change the appearance of the image to the human eye. 

Detection: The second model in SynthID is responsible for detecting this watermark. It scans the image or audio to find the watermark and confirms whether the content is AI-generated.

This dual-model approach ensures that the watermark is both effectively embedded and reliably detectable, while maintaining the original quality of the content. SynthID's design allows it to resist common forms of image manipulation, such as cropping, resizing, or filtering. This is crucial for ensuring the watermark's persistence, even when the image undergoes alterations, thereby helping to verify the authenticity of AI-generated content in various applications.

SynthID for Images 

SynthID is designed for watermarking and detecting AI-generated images, currently available to a select group of Vertex AI customers using Imagen, a cutting-edge text-to-image model.

Here's how SynthID works for images:

Watermarking: SynthID adds a digital watermark to AI-generated images (currently those generated with Imagen). It’s like a secret signature that says, "This image was made by AI." The clever part is that you can't see this watermark with your eyes—it's hidden in the image's pixels.

Detection: If someone wants to know whether an image was made by AI, they can use SynthID to check for the watermark. SynthID scans the image and tells them if the watermark is there. 

Even though SynthID puts a watermark on images, it doesn’t make the image look any different. The quality and appearance of the photo remain the same.

Moreover, the watermark stays in the image even if someone changes the picture later. This could be adding filters, changing colors, or even saving it in a different file format like JPEG, which often changes how an image looks.

Expansion to AI-generated Music and Audio

In November 2023, Google extended the capabilities of SynthID to AI-generated music and audio​​. This significant development introduced SynthID's functionality through Lyria, DeepMind's advanced AI music generation model. It means that SynthID will watermark any music or audio content published by Lyria. 

The process involves a sophisticated conversion of audio waves into a two-dimensional spectrogram, onto which the digital watermark is meticulously embedded​​. 

Audio Wave Conversion: The first step in SynthID's audio application involves converting the audio wave, a one-dimensional representation of sound, into a spectrogram. This spectrogram is a detailed visual representation that shows how the spectrum of frequencies in a sound evolves over time.

Digital Watermark Embedding: Once the spectrogram is generated, SynthID embeds a digital watermark into this visual representation. The technology is designed to integrate the watermark directly into the audio's waveform.

A key aspect of SynthID's watermark technology is to withstand common audio modifications like noise addition, compression, and alterations​​ in the track’s speed. 

Moreover, it embeds the watermark without altering the audio's perceptible qualities. This means that while the watermark is present, it remains undetectable to the human ear, thus preserving the original quality and experience of the audio.

Wrap Up 

Google's SynthID can be a game-changer in dealing with the flood of AI-generated content and deepfakes. With AI getting advanced at making content that looks and sounds real, it's tough to tell what's genuine anymore. SynthID not only helps watermark AI-made content but also helps identify AI-generated images and music.

‍