We investigate the fading cognitive multiple access wiretap channel (CMAC-WT), in which two secondary-user transmitters (STs) send secure messages to a secondary-user receiver (SR) in the presence of an eavesdropper (ED) and subject to interference threshold constraints at multiple primary-user receivers (PRs). We design linear precoders to maximize the average secrecy sum rate for multiple-input multiple-output (MIMO) fading CMAC-WT under finite-alphabet inputs and statistical channel state information (CSI) at STs. For this non-deterministic polynomial time (NP)-hard problem, we utilize an accurate approximation of the average secrecy sum rate to reduce the computational complexity, and then present a two-layer algorithm by embedding the convex-concave procedure into an outer approximation framework. The idea behind this algorithm is to reformulate the approximated average secrecy sum rate as a difference of convex functions, and then generate a sequence of simpler relaxed sets to approach the non-convex feasible set. Subsequently, we maximize the approximated average secrecy sum rate over the sequence of relaxed sets by using the convex-concave procedure. Numerical results indicate that our proposed precoding algorithm is superior to the conventional Gaussian precoding method in the medium and high signal-to-noise ratio (SNR) regimes.