Madrid, 7 (Europe Press)
They were able to develop an algorithm that effectively hides sensitive information so that it is impossible to detect that something has been hidden. The method uses new developments in information theory methods to hide one piece of content inside another so that it cannot be detected.
The team, led by the University of Oxford in close collaboration with Carnegie Mellon University, has published this work as a preliminary publication on arXiv, as well as a broken down implementation of their method on Github, and will present it at the 2023 First Artificial Intelligence Conference on Learning Learning, to be held in May.
They expect this method to be widely used soon in digital human communications, including social media and private messaging. In particular, the ability to send completely secure information can empower vulnerable groups such as dissidents, investigative journalists, and aid workers.
The algorithm applies to an environment called steganography: the practice of hiding sensitive information inside harmless content. Steganography differs from cryptography in that sensitive information is hidden in a way that obscures the fact that something has been hidden. An example of this is the disguise of a Shakespearean poem inside an AI-generated image of a cat.
Although it has been studied for more than 25 years, current steganography methods often contain imperfect security, which means that people who use these methods risk detection. This is because previous steganography algorithms subtly altered the distribution of harmless content.
To solve this problem, the team of researchers used the latest developments in information theory, specifically minimum entropy coupling, which allows two distributions of data to be linked in such a way that mutual information is maximized, but the distributions are preserved. Individually.
As a result, with the new algorithm there is no statistical difference between the distribution of harmless content and the distribution of content that encodes sensitive information.
The algorithm has been tested using several types of models that produce automatically generated content, such as GPT-2, an open source language model, and WAVE-RNN, a text-to-speech converter.
In addition to being completely secure, the new algorithm has shown up to 40% higher encryption efficiency than previous steganography methods in various applications, allowing more information to be hidden in a given amount of data. This can make steganography an attractive method even if complete security is not required, due to the advantages of data storage and compression.
The team has applied for a patent on the algorithm, but intends to license it for free to third parties for responsible, non-commercial use. This includes academic and humanitarian use and security audits from trusted third parties.
AI-generated content is increasingly being used in ordinary human communication, fueled by products like ChatGPT, Snapchat AI stickers, and TikTok video filters. As a result, steganography could become more prevalent, as the mere existence of AI-generated content would not arouse suspicion.
Dr Christian Schroeder de Witt, from Oxford University’s Department of Engineering Sciences and co-author of the study, says the method can be applied to any software that automatically generates content, such as probabilistic video filters or meme generators.
“This could be very useful, for example, for journalists and aid workers in countries where coding is illegal,” he says. However, it warns that users should continue to be careful, as any encryption technology can be vulnerable to side-channel attacks, such as detecting a steganography app on a user’s phone.
For his part, Samuel Sokota, from the Department of Machine Learning at Carnegie Mellon University, and co-author of the work, highlights that “the main contribution of this work is to show a deep connection between a problem called minimum coupling entropy and ideal secure steganography by taking advantage of this Connect, we introduce a new set of steganography algorithms that have perfect security guarantees.”
Similarly, Professor Jacob Forster, of the University of Oxford’s Department of Engineering Sciences, highlights that the study “is a wonderful example of research into the fundamentals of machine learning leading to significant advances in the field of steganography.”