Presentation
Evaluation of Multidisciplinary and Multidomain Models of Cognitive Attacks in the Mixed-Reality Domain
DescriptionFuture use of mixed reality (MR) system necessitates robust methods to ensure safe operation under adversarial conditions. DARPA’s Intrinsic Cognitive Security (ICS) program aims to address this by leveraging cognitive models and formal guarantees to mitigate attacks that may impair users’ performance. This work discusses model validation in the initial phase of ICS. The multidisciplinary nature of ICS, involving researchers from diverse fields such as psychology, machine learning, statistics, and software engineering, highlighted the necessity of establishing clear, operational definitions and guidelines early in the process. The model validation process started with the development of a rubric for assessing model fidelity, relevance, and suitability for formal methods. The iterative rubric development facilitated alignment across diverse teams, underscoring the need for clear operational definitions in multidisciplinary research while preserving the creative contributions of individual researchers. The development of a process to support model evaluation resulted in many lessons. Those lessons include: (a) It is critical to define bounds for what “validity” is considered early. (b) Construct validity is particularly challenging to agree upon, especially among multidisciplinary teams. (c) Application validity presents its own set of challenges when different teams have different experiences in the applied domain. As the ICS project progresses, these lessons (and more lessons as we encounter them) will continue to inform the development of meaningful models as the basis for formal guarantees. These guarantees, and the resulting prototypes, will effectively mitigate cognitive attacks and enhance the safe operation of MR systems in defense and other critical applications.
Event Type
Late Breaking Results
TimeTuesday, October 14th5:10pm - 5:30pm CDT
LocationGrand Hall G/H
General Sessions