Adversarial Robustness in Augmented and Virtual Reality Systems using Deep Learning Models

The rapid development of AR/VR devices led to powerful machine learning models that could withstand adversarial assaults. UUNet-based Autoencoder is adapted to increase AR/VR adversarial resilience in this study. To reduce the impact of hostile noise on efficiency, the proposed architecture uses two-stage learning to combine unsupervised feature extraction with supervised classification. Sets, including regular and malicious inputs, are learned in the UUNet-based Autoencoder. We evaluate using publicly accessible AR/VR datasets to assure input modalities and adversary variety. We evaluate the model using statistical studies, a Receiver Operating Characteristic (ROC) curve, and a confusion matrix. In the usual situation, contemporaneous empirical analysis confirms effectiveness with good accuracy, precision, recall, and F1-scores. ROC analysis showed strong discriminability and a high AUC value. When attackers noise susceptible models, detrimental perturbations cause significant performance loss. A statistical study shows that such situations' performance discrepancies are considerable because areas need improvement. The findings show that the UUNet-based Autoencoder may improve AR/VR adversarial resilience as a basic model. Adversarial training, hybrid defense, and architectural optimization should be studied to make it more reliable in AR/VR systems.