Security and Privacy for Machine Learning

Machine learning is being widely deployed in many aspects of our society. Our vision is that machine learning systems will become a new attack surface and attackers will exploit the vulnerabilities in machine learning algorithms and systems to subvert their security and privacy. In this research thrust, we aim to protect the confidentiality and integrity of machine learning algorithms and systems. Specifically, in machine learning systems, both users and model providers desire confidentiality: users desire privacy of their confidential training and testing data, while model providers desire confidentiality of their proprietary models, learning algorithms, and training data, as they represent intellectual property. We are interested in protecting confidentiality for both users and model providers. Moreover, we are interested in understanding how an attacker can compromise the integrity of machine learning systems, as well as designing new mechanisms to mitigate these attacks. In particular, for integrity, an attacker's goal is to manipulate a machine learning system such that the system makes predictions as the attacker desires. An attacker can manipulate the training phase and/or the testing phase to achieve this goal. Our ultimate goal is to build provably secure and privacy-preserving machine learning.  

Publications

Confidentiality of machine learning

Confidentiality/Intellectual Property for model providers

Confidentiality/Privacy for users

 

Integrity of machine learning

Integrity at prediction phase (i.e., adversarial examples): attacks, defenses, and their applications to privacy protection

Integrity at training phase: poisoning attacks to federated learning, graph-based methods, and recommender systems, as well as their defenses