Skip to main content

Algorithms that remember: Model inversion attacks and data protection law

Michael Veale

Many individuals are concerned about the governance of machine learning systems and the prevention of algorithmic harms. The EU’s recent General Data Protection Regulation (GDPR) has been seen as a core tool for achieving better governance of this area. While the GDPR does apply to the use of models in some limited situations, most of its provisions relate to the governance of personal data, while models have traditionally been seen as intellectual property. We present recent work from the information security literature around ‘model inversion’ and ‘membership inference’ attacks, which indicates that the process of turning training data into machine-learned systems is not one way, and demonstrate how this could lead some models to be legally classified as personal data. Taking this as a probing experiment, we explore the different rights and obligations this would trigger and their utility, and posit future directions for algorithmic governance and regulation. from Veale M, Binns R, Edwards L. forthcoming. Algorithms that remember: Model inversion attacks and data protection law. Phil. Trans. R. Soc. A 376: 20180083. (preprint:

Speaker bio

Michael Veale is an EPSRC PhD researcher in responsible machine learning at University College London, where he looks at issues of fairness, transparency and technology in the public sector, and at the intersection of data protection law and machine learning. His work on ethical and lawful use of personal data has been drawn upon by governments, Parliament and by regulators.

Share this: