Optimizing convolutional neural networks (CNNs) for real-time performance in resource-constrained settings remains a challenge. This study introduces a novel approach to CNN design, integrating compressive sensing (CS) techniques to enhance both efficiency and robustness. Our proposed CS-based CNN architecture facilitates real-time object detection in embedded environments by significantly reducing computational complexity and improving transmission efficiency. Specifically, the CS paradigm is employed as a sampling-reconstruction network within the detection CNN, resulting in streamlined processing and reduced model complexity. Previous research demonstrated a remarkable 57% reduction in floating-point operations per second (FLOPs) and a 30% increase in speed for the redesigned SSD model compared to its original counterpart [4]. In this study, we extend our investigation to assess the resiliency of the CS-based CNN architecture against adversarial perturbations. Utilizing different augmented test sets, we evaluate the performance of the CS-based model alongside its traditional baseline. Our findings reveal that the CS-based CNN exhibits superior resilience, with a smaller decline in mean average precision (mAP) under adversarial conditions: a drop of 29.1% for the CS-based model compared to a 50.8% drop for the original architecture, for the worst-case scenario. These results underscore the potential of CS-based CNNs to enable efficient and robust real-time computer vision applications in embedded environments.