Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robustness and Cybersecurity in the EU Artificial Intelligence Act

Henrik Nolte, Miriam Rateike, Michèle Finck

TL;DR

The paper investigates robustness and cybersecurity in the EU AI Act, focusing on Art. 15 high-risk AI systems and Art. 55 general-purpose AI models, and identifies definitional ambiguities that hinder practical implementation. It maps ML concepts of non-adversarial and adversarial robustness to the Act’s requirements, exploring how system-level and component-level considerations interact with accuracy and lifecycle concerns. The authors argue for harmonized standards, guidelines, and benchmarking methodologies to translate legal terms into concrete ML practices and evaluation procedures. By bridging legal terminology and ML research, the work aims to improve regulatory compliance, standardization, and the development of measurement tools for robustness and cybersecurity in AI systems.

Abstract

The EU Artificial Intelligence Act (AIA) establishes different legal principles for different types of AI systems. While prior work has sought to clarify some of these principles, little attention has been paid to robustness and cybersecurity. This paper aims to fill this gap. We identify legal challenges and shortcomings in provisions related to robustness and cybersecurity for high-risk AI systems(Art. 15 AIA) and general-purpose AI models (Art. 55 AIA). We show that robustness and cybersecurity demand resilience against performance disruptions. Furthermore, we assess potential challenges in implementing these provisions in light of recent advancements in the machine learning (ML) literature. Our analysis informs efforts to develop harmonized standards, guidelines by the European Commission, as well as benchmarks and measurement methodologies under Art. 15(2) AIA. With this, we seek to bridge the gap between legal terminology and ML research, fostering a better alignment between research and implementation efforts.

Robustness and Cybersecurity in the EU Artificial Intelligence Act

TL;DR

The paper investigates robustness and cybersecurity in the EU AI Act, focusing on Art. 15 high-risk AI systems and Art. 55 general-purpose AI models, and identifies definitional ambiguities that hinder practical implementation. It maps ML concepts of non-adversarial and adversarial robustness to the Act’s requirements, exploring how system-level and component-level considerations interact with accuracy and lifecycle concerns. The authors argue for harmonized standards, guidelines, and benchmarking methodologies to translate legal terms into concrete ML practices and evaluation procedures. By bridging legal terminology and ML research, the work aims to improve regulatory compliance, standardization, and the development of measurement tools for robustness and cybersecurity in AI systems.

Abstract

The EU Artificial Intelligence Act (AIA) establishes different legal principles for different types of AI systems. While prior work has sought to clarify some of these principles, little attention has been paid to robustness and cybersecurity. This paper aims to fill this gap. We identify legal challenges and shortcomings in provisions related to robustness and cybersecurity for high-risk AI systems(Art. 15 AIA) and general-purpose AI models (Art. 55 AIA). We show that robustness and cybersecurity demand resilience against performance disruptions. Furthermore, we assess potential challenges in implementing these provisions in light of recent advancements in the machine learning (ML) literature. Our analysis informs efforts to develop harmonized standards, guidelines by the European Commission, as well as benchmarks and measurement methodologies under Art. 15(2) AIA. With this, we seek to bridge the gap between legal terminology and ML research, fostering a better alignment between research and implementation efforts.

Paper Structure

This paper contains 35 sections, 2 figures.

Table of Contents

  1. Introduction
  2. An ML Perspective on Robustness and Cybersecurity
  3. Background on the AIA and Art. 15 AIA
  4. AIA
  5. Art. 15 AIA
  6. The Purpose of Art. 15 AIA
  7. Requirements for High-Risk AI Systems
  8. General Challenges of Art. 15 AIA
  9. Legal Requirements.
  10. Mapping to ML Understanding of Adversarial and Non-adversarial Robustness
  11. System vs. Model.
  12. Role of Accuracy.
  13. Consistent Performance Throughout the Lifecycle.
  14. Robustness Art. 15(4) AIA
  15. Inconsistent Terminology.
  16. ...and 20 more sections

Figures (2)

  • Figure 1: Robustness problems. Outputs may be as expected (✓) with "safe" test data from the original distribution; unintended changes (✗) can occur with adversarial or non-adversarial (shifted) inputs.
  • Figure 2: Technical solutions to cybersecurity can be found, i.a., in ML research on adversarial robustness; technical solutions to robustness can be found, i.a., in the ML research on non-adversarial robustness.