Challenges in Large-Scale Machine Learning Systems: Security and Correctness

Wednesday, June 12, 2019 - 2:00pm to 3:00pm
Meeting Room 2265, Innovation Center

DISSERTATION DEFENSE
Department of Computer Science and Engineering
University of South Carolina

Author : Emad Alsuwat
Advisor : Dr. Farkas and Dr. Valtorta
Date : June 12th, 2019
Time : 2:00 pm
Place : Meeting Room 2265, Innovation Center

Abstract

In this research, we address the impact of data integrity on machine learning algorithms. We study how an adversary could corrupt Bayesian network structure learning algorithms by inserting contaminated data items. We investigate the resilience of Bayesian network structure learning algorithms, namely the PC and LCD algorithms, against data poisoning attacks that aim to corrupt the learned Bayesian network model.
Data poisoning attacks are one of the most important emerging security threats against machine learning systems. These attacks aim to corrupt machine learning models by contaminating datasets in the training phase. The lack of resilience of Bayesian network structure learning algorithms against such attacks leads to inaccuracies of the learned network structure.
In this dissertation, we propose two subclasses of data poisoning attacks against Bayesian networks structure learning algorithms: (1) Model invalidation attacks when an adversary poisons the training dataset such that the Bayesian model will be invalid, and (2) Targeted change attacks when an adversary poisons the training dataset to achieve a specific change in the structure. We also define a novel measure of the strengths of links between variables in discrete Bayesian networks. We use this measure to find vulnerable sub-structure of the Bayesian network model. We use our link strength measure to find the easiest links to break and the most believable links to add to the Bayesian network model. In addition to one-step attacks, we define long-duration (multi-step) data poisoning attacks when a malicious attacker attempts to send contaminated cases over a period of time. We propose to use the distance measure between Bayesian network models and the value of data conflict to detect data poisoning attacks. We propose a 2-layered framework that detects both traditional one-step and sophisticated long-duration data poisoning attacks. Layer 1 enforces “reject on negative impacts” detection; i.e., input that changes the Bayesian network model is labeled potentially malicious. Layer 2 aims to detect long-duration attacks; i.e., observations in the incoming data that conflict with the original Bayesian model.
Our empirical results show that Bayesian networks are not robust against data poisoning attacks. However, our framework can be used to detect and mitigate such threats.