AI technologies are increasingly utilized to monitor and track individuals both in the digital realm and in real life, but these tools come with significant risks. Researchers at the Oxford Internet Institute, Imperial College London, and UCLouvain have created a new mathematical model that could enhance our comprehension of the risks associated with AI and support regulators in safeguarding personal privacy. This research has been shared in *Nature Communications*.
AI technologies are increasingly utilized to monitor and track individuals both online and in real life, but the effectiveness of these tools is accompanied by substantial risks. Researchers from the Oxford Internet Institute, Imperial College London, and UCLouvain have introduced a new mathematical model that aims to improve understanding of AI-related risks while aiding regulators in ensuring personal privacy is protected. The results of this study are featured in Nature Communications.
This method offers a solid scientific framework to evaluate identification techniques, particularly when handling vast amounts of data. For example, it can be used to assess how accurately advertising code and subtle trackers can identify online users using minimal information, such as time zone or browser settings, a technique known as ‘browser fingerprinting.’
Dr. Luc Rocher, lead author and Senior Research Fellow at the Oxford Internet Institute, part of the University of Oxford, expressed, “Our method presents a novel approach to assess the re-identification risks associated with data dissemination while also allowing evaluation of modern identification methods in critical, high-stakes situations. In environments such as hospitals, humanitarian aid operations, or border control, the implications are enormous, and accurate, reliable identification is crucial.”
The method utilizes Bayesian statistics to understand how identifiable individuals are on a small scale, and it can predict the identification accuracy for larger populations with up to ten times the effectiveness of previous strategies. This allows for an enhanced assessment of how various data identification techniques function at scale across different contexts and behaviors. Consequently, this could clarify why some AI identification methods show high accuracy in small studies but fail to correctly identify individuals in real-world scenarios.
The timing of these findings is critical, as the rise of AI-driven identification techniques presents significant challenges to anonymity and privacy. AI tools, for example, are being tested for automatic identification of individuals through their voice in online banking, their eyes in humanitarian efforts, or their face in law enforcement activities.
The researchers opine that this new model can aid organizations in balancing the advantages of AI technologies with the necessity to safeguard individuals’ personal information, thereby making interactions with technology safer and more secure. Their method identifies potential vulnerabilities and areas for refinement prior to large-scale deployment, which is vital for ensuring safety and correctness.
Co-author Associate Professor Yves-Alexandre de Montjoye from the Data Science Institute at Imperial College London stated: “Our newly developed scaling law offers, for the first time, a well-founded mathematical framework to assess how identification techniques will function at scale. Grasping the scalability of identification is critical for evaluating the risks associated with these re-identification techniques, including adherence to contemporary data protection regulations globally.”
Dr. Luc Rocher concluded by stating, “We believe our research constitutes a vital advancement toward the establishment of principled methods to evaluate the risks presented by increasingly sophisticated AI techniques and the nature of identifying human traces online. We anticipate that this work will greatly assist researchers, data protection officials, ethics committees, and other practitioners who aim to balance the sharing of data for research purposes with the necessity to protect the privacy of patients, participants, and citizens.”
