Deepfakes, Phrenology, Surveillance, and More! A Taxonomy of AI Privacy Risks

Abstract: Privacy is a key principle for developing ethical AI technologies, but how does including AI technologies in products and services change privacy risks? We constructed a taxonomy of AI privacy risks by analyzing 321 documented AI privacy incidents. We codified how the unique capabilities and requirements of AI technologies described in those incidents generated new privacy risks, exacerbated known ones, or otherwise did not meaningfully alter the risk. We present 12 high-level privacy risks that AI technologies either newly created (e.g., exposure risks from deepfake pornography) or exacerbated (e.g., surveillance risks from collecting training data). One upshot of our work is that incorporating AI technologies into a product can alter the privacy risks it entails. Yet, current approaches to privacy-preserving AI/ML (e.g., federated learning, differential privacy, checklists) only address a subset of the privacy risks arising from the capabilities and data requirements of AI.

• The paper proposes a comprehensive taxonomy of AI privacy risks, analyzing 321 incidents to understand how AI/ML alters existing privacy threats and introduces novel ones.
• Key findings include how AI exacerbates surveillance and identification risks and creates new risks like physiognomy/phrenology and exposure through generative AI (deepfakes).
• The taxonomy provides actionable insights for designing AI privacy-preserving systems, highlighting that current methods often overlook AI-specific risks and require new approaches and practitioner awareness.

A Taxonomy of AI Privacy Risks: An Overview

In the paper "Deepfakes, Phrenology, Surveillance, and More! A Taxonomy of AI Privacy Risks" by Lee et al., the authors propose a comprehensive taxonomy that seeks to understand how modern advances in AI and Machine Learning (ML) change the landscape of privacy risks. Their work is founded on an analysis of 321 documented AI privacy incidents sourced from the AI, Algorithmic, and Automation Incident and Controversy (AIAAIC) repository. With the backdrop of Solove's well-known privacy taxonomy from 2006, this paper aims to articulate the novel privacy risks introduced by AI and how existing privacy threats have been exacerbated.

The taxonomy defines twelve high-level privacy risks that can either be newly created or intensified by AI technologies. These include risks such as identification through low-quality data and the resurgence of physiognomy, where AI erroneously associates physical attributes with personal traits. The analysis reveals that AI-specific capabilities and requirements frequently alter privacy risks and provocatively argues that traditional privacy-preserving methods like federated learning and differential privacy overlook several unique threats posed by AI systems.

Key Findings

1. Data Collection and Processing Risk Dimension:
   • The research identifies Surveillance as a key risk exacerbated by AI's ability to facilitate large-scale data aggregation across diverse sources. It shows how AI systems collect vast amounts of personal data to enhance model performance, elevating the concealed gathering and analysis.
   • New processing risks arise with AI's ability to robustly identify individuals and deduce future behaviors, often from low-quality or incomplete datasets. This ability poses significant risks in various sectors, including law enforcement and personalized marketing.
2. Creation of Novel Privacy Risks:
   • AI creates an entirely new risk category labeled as Phrenology/Physiognomy, where it inadvertently promotes debunked pseudosciences by attempting to infer personality traits and demographic characteristics like criminality or sexual orientation from physical appearance alone.
   • The paper highlights exposure and distortion risks where generative AI technologies produce realistic yet fake images or videos (e.g., deepfakes), threatening personal privacy by generating non-consensual content.
3. Data Dissemination and Invasion:
   • AI exacerbates Disclosure risks through enhanced inferential capabilities, making it easier to predict sensitive individual data, as seen in contexts like China's Safe City projects for public surveillance.
   • With Intrusion, AI extends the reach of invasive technologies, turning ubiquitous devices into constant surveillance tools, hence disrupting personal solitude beyond traditional means.

Implications and Future Directions

The implications of this work span theoretical and practical domains. Practically, the taxonomy offers actionable insights for the design of AI privacy-preserving systems by demonstrating that many current privacy-protective measures only address a subset of AI-induced risks. Future AI development must account for this broader set of risks, requiring novel methodological advancements tailored to the intricate dynamics of AI technologies.

Theoretically, this study invites further exploration of privacy risks in AI-driven systems that may not have been documented yet but could emerge as AI continues to infiltrate diverse sectors. Potential future risks include interrogation through AI-driven conversation tools, breaches of trust with AI mediating confidential interactions, and AI's role in enhancing or inducing new forms of decisional interference.

As the landscape of AI capabilities evolves, the taxonomy is seen as a living document that researchers and practitioners must iteratively refine in tandem with emerging AI incidents. The study underscores the need for enhanced education and awareness among AI practitioners regarding the holistic perspective on privacy that considers AI-specific risks.

In conclusion, by articulating how AI changes the privacy risk paradigm, Lee et al. provide a critical foundation for AI researchers and practitioners to both anticipate and address the unique challenges introduced by integrating AI into everyday applications. This taxonomy serves as a pivotal resource in the ongoing endeavor to responsibly innovate in AI while safeguarding individual privacy.