Page 37 - 陽明交大電機學院院刊 2025.02

同時，本論文提出了一種個人化隱私 型之間的差距，並將平均測試準確率提升
保護聯邦隨機梯度下降算法 (P- 了 3.82%。與傳統聯邦學習算法相比，原

PFedSGD)，允許客戶端在訓練過程中傳 本需要 226 個回合才能達到 85%的準確

送梯度而非模型參數，從而有效解決模型 率，而 P-PFedSGD 展現了更快的收斂速
參數傳輸可能引發的隱私洩露問題。 度，僅需 38 個回合即可達到相同的準確

率，顯著降低了客戶端的通訊與計算量。
在刀具磨耗資料及的實驗結果中，P-
PFedSGD 有效縮小了全局模型與本地模

IEEE Transactions on Neural Systems and Rehabilitation

Engineering

Exploring Embodied Cognition and Brain Dynamics
Under Mul-ti-Tasks Target Detection in Im-merse
Projector-Based Aug-mented Reality (IPAR) Scenarios

2024/09,

電控工程研究所 柯立偉 教授（通訊作者）

clustering. Key findings include decreased
This study explores embodied cognition by
behavioral performance during multitasking, re-
inves-tigating how mobile technology in
duced alpha and beta power in the frontal and
immersive hy-brid settings influences cognitive
motor cortex, and notable perturbations in theta
processes. We used immersive projector-based
power during distraction tasks. Walking tasks in-
augmented reali-ty (IPAR) and wireless EEG to
duced more significant neural fluctuations than
examine human cognition in multitasking
environmental distractions, particularly in beta
scenarios with mixed body movements and
suppression. These results highlight the dynamic
environmental distractions. Fifteen participants
brain-body interaction in multitasking and the
engaged in four multitasking conditions
value of integrating immersive augmented reality
(standing/walking with and without distraction),
into embodied cognition research, offering in-
and EEG data were processed using Independent
sights for improving human-computer interaction
Component Analysis (ICA) and K-means
and understanding cognitive dynamics.

36