Robust Engagement Estimation in Mixed Reality Under Cognitive Attack

Engagement is a critical operational state in immersive systems, where users must sustain attention to virtual content despite distraction, workload, and degraded system responsiveness. In MR, one important source of disruption is a cognitive attack: an adversarial or system-level perturbation to the interaction loop that interferes with attention, situational awareness, or stable task execution. In this paper, we study induced latency as a concrete MR-relevant cognitive attack and present a two-stage model for estimating engagement from multimodal time-series data, including eye tracking, head motion, scene dynamics, and system latency. The first stage uses contrastive learning to encode variable-length behavioral trajectories into compact embeddings that suppress nuisance variation while preserving engagement-relevant structure. A lightweight regression head then maps these embeddings to a continuous engagement score. We evaluate the framework in a controlled hostage-rescue MR study with step-wise latency perturbations and minute-level engagement self-reports. Across cross-validation and held-out-user evaluation, the proposed model performs favorably relative to traditional regressors on pooled features, indicating that temporally structured multimodal representations provide a measurable benefit for engagement estimation under attack. We further show how formal verification can be used to derive conservative safe operating intervals for controllable MR parameters, connecting engagement estimation to runtime reasoning in safety-aware MR systems.