The landscape of facial processing applications in the context of the European AI Act and the development of trustworthy systems

Isabelle Hupont, Songül Tolan, Hatice Güneş, Emília GómezScientific 2022 · DOI

Abstract

This work focuses on facial processing, which refers to artificial intelligence (AI) systems that take facial images or videos as input data and perform some AI-driven processing to obtain higher-level information (e.g. a person’s identity, emotions, demographic attributes) or newly generated imagery (e.g. with modified facial attributes). Facial processing tasks, such as face detection, face identification, facial expression recognition or facial attribute manipulation, are generally studied as separate research fields and without considering a particular scenario, context of use or intended purpose. This paper studies the field of facial processing in a holistic manner. It establishes the landscape of key computational tasks, applications and industrial players in the field in order to identify the 60 most relevant applications adopted for real-world usages. These applications are analysed in the context of the new European AI Act and the 7 requirements for Trustworthy AI defined by the European High Level Expert Group on AI. More particularly, we assess the risk level conveyed by each application according to the AI Act and reflect on current research, technical and societal challenges towards trustworthy facial processing systems.

Introduction

Facial processing (FP) stands as one of the most scrutinized and rapidly evolving domains within artificial intelligence. Despite its prevalence, academic research has historically been siloed into narrow technical objectives—such as improving the accuracy of face detection or refining emotion recognition—often divorced from the actual context of use. However, the regulatory environment is shifting. The European AI Act regulates AI based on its intended purpose and associated risk level, rather than just the underlying mathematical models. This paper addresses the critical need for a holistic overview that bridges the gap between technical AI capabilities and their legal and ethical implications.

Mapping the Technical Landscape

To provide a comprehensive map of the field, the authors conducted a multi-stage analysis of a decade of research (2012–2021) and the portfolios of 183 companies. This resulted in the identification of 12 fundamental computational tasks that form the building blocks of most facial processing systems.

Overview of the 12 fundamental facial processing tasks identified in the study.

These tasks range from basic Face Detection (FD) and Tracking (FT) to more complex operations like Facial Attribute Manipulation (FAM) and Automatic Lip Reading (ALR). By mapping these tasks to 60 distinct real-world applications, the study highlights how technical components are combined to serve specific industrial needs, from airport security to social media filters.

Data Challenges: Noise and Bias

A significant portion of the paper examines the datasets that fuel these technologies. While no new dataset is introduced, the meta-analysis reveals troubling trends in existing benchmarks like MS-Celeb-1M, WebFaces42M, CelebA, and LFW.

  • Data Quality: Public datasets are often "noisy," with mislabeling rates reaching up to 50% in some instances.
  • Demographic Imbalance: Most datasets are heavily biased toward Western, white, and "conventionally attractive" individuals.

These flaws present a major hurdle for developers attempting to comply with the AI Act’s stringent requirements for "error-free" and representative training data. The authors advocate for the adoption of protocols like FRUITS to ensure more demographic-aware evaluations.

The Framework for Trustworthy AI

The core of the paper’s analysis lies in evaluating facial processing through the lens of the "7 Requirements for Trustworthy AI" and the risk tiers of the EU AI Act.

The conceptual framework linking facial processing applications and tasks to the seven requirements for Trustworthy AI.

The authors manually labeled 60 applications according to four risk levels defined in the Act: Unacceptable, High-risk, Transparency-risk, and Minimal/No-risk. This categorization helps stakeholders understand the regulatory burden associated with specific use cases.

Key Insights and Market Dynamics

The study reveals that the risk profile of an algorithm is highly dependent on its context. For instance, face identification used for smartphone authentication is considered minimal risk, whereas the same technology used for public surveillance by authorities is strictly regulated.

Furthermore, the market analysis shows that 66% of companies in this space are Small and Medium Enterprises (SMEs). These smaller players face a significant disadvantage, as they often lack the resources to acquire the high-quality, audited datasets necessary for regulatory compliance. The authors also warn against the "accuracy trap," noting that a system with high aggregate accuracy may still exhibit discriminatory failures against minority groups.

Conclusion and Future Directions

The transition toward trustworthy facial processing requires moving beyond "black box" deep learning models toward systems that offer explainability and robust human oversight. The authors suggest that "Datasheets for Datasets" should become an industry standard to ensure auditable data governance. Additionally, privacy-preserving techniques like Federated Learning represent a promising, though currently under-utilized, path forward. As the legal text of the AI Act continues to be finalized, this landscape provides a vital foundation for developers and policymakers to navigate the ethical complexities of facial processing.

Summary updated

Abstract

This work focuses on facial processing, which refers to artificial intelligence (AI) systems that take facial images or videos as input data and perform some AI-driven processing to obtain higher-level information (e.g. a person's identity, emotions, demographic attributes) or newly generated imagery (e.g. with modified facial attributes). Facial processing tasks, such as face detection, face identification, facial expression recognition or facial attribute manipulation, are generally studied as separate research fields and without considering a particular scenario, context of use or intended purpose. This paper studies the field of facial processing in a holistic manner. It establishes the landscape of key computational tasks, applications and industrial players in the field in order to identify the 60 most relevant applications adopted for real-world uses. These applications are analysed in the context of the new proposal of the European Commission for harmonised rules on AI (the AI Act) and the 7 requirements for Trustworthy AI defined by the European High Level Expert Group on AI. More particularly, we assess the risk level conveyed by each application according to the AI Act and reflect on current research, technical and societal challenges towards trustworthy facial processing systems.
BibTeX
@inproceedings{hupont2022landscape,
  author = {Isabelle Hupont and Songül Tolan and Hatice Güneş and Emília Gómez},
  title = {{The landscape of facial processing applications in the context of the European AI Act and the development of trustworthy systems}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-14981-6},
}
Figure