[Paper Review] Federated Learning Priorities Under the European Union Artificial Intelligence Act

Herbert Woisetschläger, Alexander Erben+5 more|arXiv (Cornell University)|Feb 5, 2024

Digital Transformation in Law9 citations

TL;DR

The paper analyzes how the EU AI Act affects Federated Learning (FL) and argues for reframing FL research priorities to meet data governance, privacy, energy efficiency, and robustness requirements under the Act.

ABSTRACT

The age of AI regulation is upon us, with the European Union Artificial Intelligence Act (AI Act) leading the way. Our key inquiry is how this will affect Federated Learning (FL), whose starting point of prioritizing data privacy while performing ML fundamentally differs from that of centralized learning. We believe the AI Act and future regulations could be the missing catalyst that pushes FL toward mainstream adoption. However, this can only occur if the FL community reprioritizes its research focus. In our position paper, we perform a first-of-its-kind interdisciplinary analysis (legal and ML) of the impact the AI Act may have on FL and make a series of observations supporting our primary position through quantitative and qualitative analysis. We explore data governance issues and the concern for privacy. We establish new challenges regarding performance and energy efficiency within lifecycle monitoring. Taken together, our analysis suggests there is a sizable opportunity for FL to become a crucial component of AI Act-compliant ML systems and for the new regulation to drive the adoption of FL techniques in general. Most noteworthy are the opportunities to defend against data bias and enhance private and secure computation

Motivation & Objective

• Analyze the impact of the EU AI Act on Federated Learning (FL) systems and identify regulatory-aligned research priorities.
• Evaluate data governance, privacy, energy efficiency, and robustness requirements for FL under the AI Act.
• Provide quantitative and qualitative insights on costs, trade-offs, and opportunities for FL compliance and adoption.

Proposed method

• Perform a data governance and privacy-centric analysis aligned with AI Act priorities.
• Quantitatively assess energy and computational costs of privacy-preserving techniques in FL (e.g., DP, SMPC, HEC).
• Qualitatively analyze data lineage, bias mitigation, and governance advantages of FL under GDPR and the AI Act.
• Use a BERT fine-tuning FL experiment to quantify privacy-energy trade-offs and validation costs.
• Propose future research priorities to accelerate FL adoption in compliant ML systems.

Experimental results

Research questions

• RQ1 How does the EU AI Act influence Federated Learning in terms of data governance, privacy, energy efficiency, and robustness?
• RQ2 What are the trade-offs and costs of applying private and secure computation techniques in FL under AI Act requirements?
• RQ3 Can FL's data lineage and siloed-data access help address data bias and GDPR-like privacy requirements in high-risk applications?
• RQ4 What are practical research priorities to align FL with the AI Act for broad adoption?

Key findings

• The AI Act introduces governance, privacy, energy, and robustness challenges that may shift FL research priorities.
• Privacy techniques differ in cost: DP is lighter on computation but complicates regulatory alignment, while SMPC/HEC are more costly but offer strong privacy.
• Energy efficiency is a critical constraint; FL currently trails centralized training in energy efficiency, with PEFT showing substantial energy savings.
• Validation and monitoring under the AI Act incur energy and scheduling costs, creating an optimization problem between timely robustness checks and energy use.
• FL’s privacy-by-design nature and data lineage advantages position it as a favorable framework for compliant high-risk applications, given appropriate research focus.
• The paper provides a quantified privacy-energy trade-off and outlines qualitative advantages of FL in meeting GDPR/AI Act privacy requirements.