Swarm Characteristic Classification using Robust Neural Networks with Optimized Controllable Inputs

TL;DR

This work addresses robust tactic classification of attacking swarms under operational uncertainty by enriching training data with controllable and uncontrollable variables (notably defender number $N_D$, defender motion $DM$, and measurement noise) and by developing an optimal defender-trajectory framework to maximize the neural network's Sum of True Predictions $STP$. The approach combines a CNN-based time-series classifier with an optimal-control formulation that generates defender trajectories respecting dynamical, area, and collision constraints, aiming to minimize defender resources while maximizing classification confidence. The key contributions include systematic VOI selection for data augmentation, comprehensive robustness analyses across defender count, motion, and noise, and a deployment framework that yields actionable defender trajectories and minimum defender requirements. The results demonstrate that training on enriched, combined datasets substantially improves robustness and that optimized defender motions can achieve near-maximum $STP$, enabling efficient and adaptable swarm-defense planning for both military and civilian autonomous-systems contexts.

Abstract

Having the ability to infer characteristics of autonomous agents would profoundly revolutionize defense, security, and civil applications. Our previous work was the first to demonstrate that supervised neural network time series classification (NN TSC) could rapidly predict the tactics of swarming autonomous agents in military contexts, providing intelligence to inform counter-maneuvers. However, most autonomous interactions, especially military engagements, are fraught with uncertainty, raising questions about the practicality of using a pretrained classifier. This article addresses that challenge by leveraging expected operational variations to construct a richer dataset, resulting in a more robust NN with improved inference performance in scenarios characterized by significant uncertainties. Specifically, diverse datasets are created by simulating variations in defender numbers, defender motions, and measurement noise levels. Key findings indicate that robust NNs trained on an enriched dataset exhibit enhanced classification accuracy and offer operational flexibility, such as reducing resources required and offering adherence to trajectory constraints. Furthermore, we present a new framework for optimally deploying a trained NN by the defenders. The framework involves optimizing defender trajectories that elicit adversary responses that maximize the probability of correct NN tactic classification while also satisfying operational constraints imposed on the defenders.

Figures (16)

• Figure 1: Different adversary responses when varying defender number, even with identical engagement initialization. Velocity vector shown for each agent.
• Figure 2: Five distinct defender motions elicit unique adversary responses, even with identical initialization.
• Figure 3: Overview of the trajectory optimization process for one defender resulting in an optimized response from the adversary regardless of tactic employed.
• Figure 4: Defender Number Experiment: comparison of 5 NN; four NN trained with fixed defender number ($N_D$), and one NN trained using "Combined" dataset ($N_D=[1:15]$).
• Figure 5: Accuracy improves slightly when a NN is trained with more defenders than adversaries.