Leveraging Adversarial Machine Learning Techniques for Deceptive Sampling-Based Motion Planning
Author(s)
Nichols, Hayden
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
There are many applications in which a mobile agent wants to avoid having its intent known to an observer. Additionally, a mobile agent may want to have deceptive actions that convey an intent other than its true objective. Examples of this include preserving privacy in a high-surveillance environment or confusing an opponent in an adversarial setting. However, this desire for deception can conflict with the need for an efficient path. Optimal plans such as those produced by RRT* may have low path cost. However, optimality can lead to predictability, in that observers can often predict the intent of an optimal agent. Similarly, a deceptive path that moves in such a way to confuse the observer may take too long to reach the goal. This work attempts address this trade-off by drawing inspiration from adversarial machine learning. Presented in this thesis is a novel planning algorithm, dubbed Adversarial RRT*. Adversarial RRT* attempts to deceive machine learning classifiers by incorporating a predicted measure of deception into the planner cost function. Adversarial RRT* considers both path cost and a measure of predicted deceptiveness in order to produce a trajectory with low path cost that still has deceptive properties. Performance of Adversarial RRT* is demonstrated with two measures of deception, using a simulated Dubins vehicle. Adversarial RRT* can decrease RNN accuracy across paths to 10%, compared to 46% accuracy on near-optimal RRT* paths, while keeping path length within 16% of optimal. This work also presents two simulated use cases of Adversarial RRT* planner. In one use-case, the planner attempts to safely deliver a high value asset to a predetermined location while an adversary observes the vehicle's path and tries to interrupt the delivery. In the other use-case, fixed-wing delivery drone dynamics are approximated using a Dubins' vehicle model and Adversarial RRT* is demonstrated as a method to preserve the privacy of the recipient of an aerial delivery.
Sponsor
Date
2022-05-03
Extent
Resource Type
Text
Resource Subtype
Thesis