1: \begin{abstract}

2: In this paper, we investigate  secure transmission over the large-scale

3: multiple-antenna wiretap channel with finite alphabet inputs.

4: First, we show analytically that a generalized singular value decomposition (GSVD)

5: based design, which is optimal for Gaussian inputs,

6: may exhibit a severe performance loss for finite alphabet inputs in the high signal-to-noise ratio (SNR) regime.

7: In light of this, we propose a novel Per-Group-GSVD (PG-GSVD) design

8: which can effectively compensate the performance loss caused by the GSVD

9: design. More importantly,  the computational complexity of the PG-GSVD design  is

10:  by orders of magnitude lower than that of the existing

11: design for finite alphabet inputs in \cite{Wu2012TVT} while the resulting performance loss is minimal.

12: Numerical results  indicate that the proposed PG-GSVD design can be efficiently implemented

13: in large-scale multiple-antenna systems and achieves significant performance gains compared to the GSVD design.

14: \end{abstract}

15: