Mild cognitive impairment (MCI) detection using eye-tracking (ET) signals in virtual reality (VR)-based cognitive tasks shows great promise, as it can capture rich temporal and behavioral information. Therefore, we build four VR-based tasks based on Stroop test and construct a dataset for MCI detection using ET signals. However, ET signals often suffer from non-stationarity,variability, and redundancy, challenging accurate MCI detection.To address these issues, we propose a novel lightweight network KWS-TA-CNN with three key components: 1) Kymatio Wavelet scattering transform (KWS), which generates time-robust features and reduces memory usage through a depth-first traversal strategy; 2) Temporal Attention (TA) to dynamically weight critical time steps for MCI detection; and 3) 1D Convolutional Neural Network (CNN) to capture local temporal patterns and reduce feature redundancy. Experimental results from leave-one-subject-out cross-validation show high performance, with subject-level accuracies of 0.8158, 0.9211, 0.8158, and 0.8421 across the four tasks, demonstrating its strong clinical potential.