革新知能統合研究センター

In recent years the areas of machine learning (ML) and security have
both received tremendous attention from the research community.
Interestingly these two areas mutually benefit from each other. Recent
advances in machine learning, and especially in deep learning, have
enabled new security capabilities. Conversely, ML algorithms are major
targets for attackers who intend to compromise the model security and
data privacy. Motivated by the security and privacy challenges of health
data access for deep learning predictive models, in this talk I will
discuss the importance of conducting research at the intersection of
security and ML, and describe a few projects that we are currently
working on to address confidentiality (optimization-based anonymization)
and integrity (defence against poisoning attacks) of machine learning
models using optimization techniques (MILP) and game theory. I will also
discuss why security and privacy of machine learning models need to be
addressed beyond the classical security model of confidentiality,
integrity and availability and report on an accountability framework for
ML models.