Abstract
Human-Machine Teaming (HMT) systems, comprising humans, AI agents, and autonomous vehicles, are increasingly deployed in high-stakes environments such as search and rescue and surveillance. These teams rely on a Shared Mental Model (SMM) for effective coordination; however, traditional SMMs assume all teammates are inherently trustworthy. This paper introduces DASH (Deception-Augmented Shared Mental Model), a novel framework that secures HMTs by embedding proactive "bait tasks" into their coordination logic. By strategically testing teammates with tasks that have known outcomes, DASH can detect and replace compromised agents in real-time. Our evaluation shows that DASH maintains an 80% mission success rate under heavy adversarial attacks—an eightfold improvement over baseline systems—while significantly reducing the rate of compromised team members.
Introduction
In modern tactical environments, the synergy between humans, AI agents, and Unmanned Ground Vehicles (UGVs) is critical. Effective collaboration is facilitated by a Shared Mental Model (SMM), which provides a collective understanding of tasks, roles, and mission goals. Despite their utility, SMMs are vulnerable to "insider threats." A UGV’s communication link may be hijacked, an AI model might harbor a hidden backdoor, or a human analyst could be compromised via social engineering.
Traditional security measures, such as firewalls and encryption, often fail to detect a teammate that has already bypassed perimeter defenses and is acting maliciously from within. DASH addresses this critical gap by shifting from a model of implicit trust to one of continuous, proactive verification. By integrating cyber deception directly into the team's coordination framework, DASH ensures mission resilience even when individual components can no longer be trusted.
The DASH Framework
DASH integrates cyber deception into the HMT coordination logic through several key components designed to "smoke out" compromised entities.
Adaptive Deceptive Task Management (ADTM)
The ADTM serves as the framework's core logic for task allocation. It dynamically intersperses regular mission objectives with "bait tasks"—assignments where the outcome is already known to the command center.
- For UGVs: The system may instruct a vehicle to "detect" an object strategically placed in a known location.
- For AI Agents: The system provides images with pre-validated labels to check for backdoor triggers or adversarial perturbations.
- For Human Analysts: The system presents previously verified imagery to identify biased or incorrect reporting.
Dynamic Trust Management
Every teammate is assigned a trust score ($T_X$) that evolves based on their performance on both mission-critical and bait tasks. This scoring system includes a time-based decay factor, requiring teammates to consistently demonstrate reliability to maintain their standing within the SMM. If a teammate's trust score falls below a predefined threshold ($\zeta$), DASH triggers specific recovery protocols, such as software reinstallation for UGVs, model retraining for AI, or personnel replacement for humans.
Evidential Deep Learning (EDL)
To handle the inherent uncertainty in AI decision-making, DASH utilizes Evidential Deep Learning and Subjective Logic. This allows the AI agent to quantify its own uncertainty (vacuity). When the AI identifies high uncertainty in its own processing, it proactively triggers a request for human intervention, preventing the propagation of unreliable data through the SMM.
Experimental Evaluation
The DASH framework was evaluated using a simulated surveillance mission. The AI agent utilized a ViT-B/16 (Vision Transformer) pre-trained on ImageNet and fine-tuned on a subset of CIFAR-10 (6,000 images) representing common surveillance targets like vehicles, humans, and animals.
The performance was measured across 200 simulated object detection missions, testing the system against varying attack rates and member vulnerabilities. DASH was compared against three alternative schemes: an SMM-only model, a Defense-only model (without SMM integration), and a baseline system with no specialized security logic.
Results and Discussion
The experimental results highlight the robust resilience provided by the DASH framework:
- Mission Success: Under high-intensity adversarial attacks, DASH sustained an 80% mission success rate. In contrast, the baseline system collapsed to below 10% success under the same conditions.
- Compromise Mitigation: DASH reduced the presence of compromised members within the team by up to 70% compared to standard coordination approaches.
- Information Integrity: The SMM Quality Index (SQI), which measures the accuracy of shared information, remained stable at approximately 70% for DASH. Other models saw their shared data become almost entirely unreliable as attacks progressed.
While DASH introduces an operational overhead—approximately 20–30% higher cost due to the execution of bait tasks and recovery protocols—this "security tax" is essential for preventing total mission failure in contested environments.
Conclusion
DASH represents a paradigm shift in Human-Machine Teaming security, moving from reactive defense to proactive verification through deception. By utilizing internal deception as a diagnostic tool, the framework maintains high mission success rates and information integrity even when teammates are compromised.
Future work will focus on transitioning DASH from simulation to physical testbeds involving real UGVs and human operators. Additionally, researchers aim to explore adaptive adversaries who may attempt to learn the patterns of bait tasks, as well as the long-term psychological effects on human operators who are subject to continuous deceptive testing.
Summary updated