[Hugging Face] - The Battle Against Deepfakes through AI Watermarking
Watermarking in Images, Text and Audio
The Synopsis:
Hugging Face, an open-source machine learning company, has outlined different, complex techniques to deter the spread of deepfakes through AI Watermarking. AI Watermarking is used “ to mark content in order to convey additional information, such as authenticity.” These watermarks can be visible and recognizable to humans and/or machines, and there are two types of watermarking techniques: during or after the AI-generated content.
The two methods have their strengths; before content creation, the watermarking is more robust and automatically embedded into content, and after the content creation, the watermark can be applied to closed-sourced and proprietary content. Additionally, watermarks can be closed or open, yet one or the other poses risks to usability and security, respectively; a hybrid of open-closed watermark, like the one used in Truepic would address both risks.
In addition to watermarking, data poisoning and signing techniques are also used to deter deepfake images. Software for data ‘poisoning’, such as Glaze, Photoguard, Nightshade, and Fawkes imperceptibly alter images so algorithms cannot process them well. Truepic uses a C2PA standard that validates the metadata and integrates watermarking to mitigate the removal of information.
Watermarking is different for images, text, and audio, and there are different mechanisms to ensure the safety of content online:
Although the technology is novel, there is promise that these technologies will mitigate algorithmic harms.1
The Analysis:
Hugging Face has researched the various watermarking methods to deter deepfakes, and from this research, different companies can implement into their AI-generated content. However, I am unsure of the scaling of these methods. For example, Nightshade appears to be a promising deterrent to deepfake images, but I suspect that for-profit companies benefit from manipulating seemingly innocent images. Recently, I have seen an original photo of Audrey Hepburn become an AI-generated video, in which she is singing Perfect by Ed Sheeran. If there are companies that profit from manipulating images and videos to sell a product, then I question their incentives to act responsibly, ethically with technology. Seemingly, the ‘utilization’ from these Generative AI technologies is contorting original, authentic content and repurposing it for profit, regardless of its ethical implications. However, I believe the Generative AI models can be used more creatively, making authentic, original AI content, yet companies have relegated to copying, manipulating, and pasting. Generative AI is a tool or weapon, depending on who is using it. Unfortunately, Generative AI has been weaponized, even in small ways, and AI Ethicists, like myself, have to continue addressing these ‘use cases’ to mitigate these algorithmic harms.
Ethical AI is falling behind the ever-growing, unfettered technologies, and the risks are increasing, such as the incident of the deepfake call of Joe Biden — The US President. We are heading into election season, which will reveal the worst of humanity because politicians, companies, citizens all have juxtaposed values. Artificial Intelligence would only make racism, sexism, climate insensitivity efficient, and I inquire to myself is this our fate as a society. Rhetorically, can we withstand an overwhelming wave of disinformation on which we base our truths? Are future generations going to inherit the calamity that we had the ability to shield them from? These are questions that leaders of organizations and governments need to ask themselves and on which they need to reflect. If they are unwilling and unable to safeguard employees and citizens from extensive algorithmic harms, then I suggest they be removed from their position and replaced with people who care about a human-centered technology society. History accounts for those who have and have not stood for justice, fairness, and humanness, and Hugging Face is settling into its place as proponent for an open-source, algorithmically safe organization. We have a history to tell our grandchildren that we can truthfully and happily say to them that we have fought for their future, and it us up to them to continue carrying the torch for benevolent global change.
The Endnotes:
1 Sasha Luccioni, Yacine Jernite, Derek Thomas, et al.,“AI Watermarking 101: Tools and Techniques,“ Hugging Face, accessed Mar 21, 2024,