Algorithms that remember: model inversion attacks and data protection law.

Area of Science:

• Artificial Intelligence Law
• Information Security
• Data Governance

Background:

• Growing concerns exist regarding the governance of machine learning (ML) systems and the prevention of algorithmic harms.
• The EU's General Data Protection Regulation (GDPR) is a key regulatory framework, but its application to ML models is complex.
• Traditionally, ML models are viewed as intellectual property, distinct from personal data.

Purpose of the Study:

• To investigate whether ML models can be legally classified as personal data under GDPR.
• To explore the implications of such a classification for algorithmic governance and regulation.
• To analyze the utility of GDPR rights and obligations in the context of ML models.

Main Methods:

• Review of information security literature on 'model inversion' and 'membership inference' attacks.
• Analysis of the legal implications of these attacks on the classification of ML models.
• Exploration of potential GDPR rights and obligations triggered by models being classified as personal data.

Main Results:

• Model inversion and membership inference attacks demonstrate that the process of creating ML models from data is not unidirectional.
• These attacks suggest that some ML models could be legally interpreted as containing or representing personal data.
• This potential classification necessitates a re-evaluation of current algorithmic governance frameworks.

Conclusions:

• The findings challenge the traditional view of ML models as solely intellectual property.
• Classifying models as personal data under GDPR could introduce new avenues for algorithmic governance and user rights.
• Further research is needed to explore future directions for AI governance and regulation based on these insights.