Privacy and Security in Machine Unlearning

This chapter turns to the privacy and security implications that arise when machine unlearning moves from algorithmic ideal to deployed practice. While unlearning is often motivated by data protection regulations and the desire to mitigate memorisation of sensitive information, unlearning procedures are vulnerable to being utilized by adversaries, leaving residual traces that are exploitable by inference and reconstruction attacks, or opening up new avenues for model manipulation by malicious users or servers.