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 investigate the case where instantaneous channel state information (CSI)

5: of the eavesdropper is known at the transmitter. We show analytically that a generalized singular value decomposition (GSVD)

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

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

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

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

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

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

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

13: Then, we extend the PG-GSVD design to the case where only statistical CSI of the eavesdropper

14: is available at the transmitter.

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

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

17: \end{abstract}

18: