Automated Modeling of Deceptive Intent in Computer-Mediated Conversations
Abstract
Social interactions are rife with uncertainties as to the intent and actions of adversaries. While understanding deception in interactions is an incredibly hard problem (even confounding humans), use of automated methods could avoid human biases. The proposed work will collect data on several hundred dyadic interactions, and analyze them for non-verbal cues that suggest deceptive practices. In traditional Machine Learning style work, part of the data will be used for training and part of the data will be used for validation. It is also expected that data collected at ARDEC might also be used for validation. The proposed approach, pioneered by the PI at MIT Media Lab, is based on using non-verbal cues, including prosody, gestures, body language, etc, to understand the mental state of a person involved in an interaction. The proposed work will be based on automatically detecting inconsistencies in feature space of each participant in an interaction, where the feature space will include language spoken, gestures, prosody, etc. Indeed, the proposed work will build upon, and test, theories of deception being formulated in the social sciences community. By carrying out extensive data collection, which will be used for both training and for validation, the PI will ensure that the proposed work is grounded in reality. Finally, researchers at ARDEC have shown interest in collaborating with the PI (if the work is funded), which the PI will use for additional validation of his work.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 12, 2017
- Source ID
- W911NF1510157
Entities
People
- Mohammed Hoque
Organizations
- Army Contracting Command
- United States Army
- University of Rochester