10.3.3 Adversarial Reasoning: Inferring Trustworthiness and Deceit with Adversarial Relational Models
Abstract
It is often difficult to distinguish honest information and agents from deceitful ones. This is a broad problem with far-reaching applications, including deciding which informants to trust, deciding which pieces of intelligence are reliable, estimating the risk of insider threats, and more. Trust is an inherently relational attribute: we trust information and people based on their relationships to other trusted entities, as well as their own individual attributes. These domains are also adversarial: deceitful entities will attempt to disguise their behavior in order to appear trustworthy, modifying both their behavior and their relationships. Therefore, any effective system must incorporate both of these aspects simultaneously. The goal of this project is to develop robust methods for detecting deceit in realistic, large-scale relational domains. This will be accomplished through three aims. Aim 1: Apply existing adversarial relational methods to detecting trustworthiness and deceitfulness in two representative real-world domains: fake product reviews and comment spam. Fake product reviews are difficult to distinguish from real ones, but the reviewer-review-product graph shows structural signatures for fraudulent reviews. For comment spam, graph motifs among users and comments have been used to identify spam campaigns in YouTube. In this project, we will build these features and others into a relational model that can reason jointly about all comments and users, and is robust to attempts to disguise these signatures. Aim 2: Develop scalable inference and learning methods for adversarial problems. We will build on state-of-the-art methods from statistical relational learning, extending them to handle adversarial settings. Adversarial problems are likely to feature certain kinds of structure, such as many actions that are roughly equivalent. We will exploit these symmetries to reduce the action space exponentially. Aim 3: Extend adversarial relational learning methods to the general-sum setting, in which the learner and adversary have goals that are not directly opposed to each other. By modeling these utilities more accurately, we will be able to learn more effective models while remaining robust to adversarial manipulation.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2018
- Source ID
- W911NF1510265
Entities
People
- Daniel Lowd
Organizations
- Army Contracting Command
- United States Army
- University of Oregon