Although the initial concept about secrecy transmission can be traced back to the 1970s , wireless transmission issues concerning physical layer security rate have attracted considerable attention in recent years. Traditional communication methods based on High-level encryption can hardly be used to improve physical layer security rate in practical situation, such as WLAN and Ad-hoc network. In WLAN scenario, the unpredictable random access and leave of users would lead to difficulties for the establishment of an appropriate and reasonable high-level encryption protocol. In addition, in Ad-hoc networks, a complete data transmission could go through several hops and other users may participate in relaying data, which would decrease secrecy rate.
One basic idea on the physical layer security is artificial noise. It adds artificial noise to the transmission signal expecting that the artificial noise would provide more negative effects for eavesdroppers than legitimate users. In  and , the authors analyze the effect of artificial noise on enhancing secrecy capacity. The difference is that in  the transmitter generates the artificial noise, while in  an external node helps to accomplish this work. In paper , the secure transmission in multiuser and multiple-eavesdropper systems is investigated. Several strategies are illustrated to suppress the channel interference. Numerical results demonstrate that adding artificial noise to the transmission message would improve the system secrecy rate. In paper , the authors investigate secure communication between two multi-antenna nodes with an undetected eavesdropper. Unlike previous work, no information regarding the eavesdropper is available. Artificial interference is applied to mask the desired signal. The authors maximize the power available to hide the desired signal from a potential eavesdropper, while maintaining a prespecified signal-to-interference-plus-noise-ratio (SINR) at the desired receiver. The case of the presence of imperfect channel state information (CSI) is also studied.
Another category to protect the physical layer security is based on beamforming. In , a relay assisted system is studied with the assumption that the relay is unreliable. Precoding designs of the base station (BS) and relay are provided to maximize the secrecy capacity. In addition,  takes the two-way relay scenario into consideration. In this situation, eavesdroppers would utilize the received information obtained from two transmission slots to get the desired eavesdropped users’ information. Optimal precoding design strategy in this scenario is presented. In , a joint beamforming design of the source and relay based on quality-of-service (QoS) requirements with presence of channel uncertainty is investigated.  studies the secrecy rate in decode-and-forward relay scenario with finite-alphabet input. A power control scheme based on semidefinite programming is presented for the purpose of maximizing secrecy rate with finite-alphabet input.  investigates the relay-eavesdropper network with two models of imperfect knowledge of the eavesdropper s channel. The approximation of the ergodic secrecy rate is studied under the Rician fading channel model and the worst-case secrecy rate is considered under the deterministic uncertainty model. Under both models, the optimal rank-1, match-and-forward (MF), and zero-forcing (ZF) beamformers are developed. The effectiveness of the proposed relay beamformers are verified by the numerical results. Paper  provides precoding strategies in a coordinated multi-point (CoMP) transmission system.
Robust design of beamforming and artificial noise has been investigated in multiple-input-single-output (MISO) networks. In , the authors address the physical layer security in MISO communication systems. The transmission covariance matrices of the steering information and the artificial noise are investigated to maximize the worst-case secrecy rate in a resource-constrained system and to minimize the use of resources to ensure an average secrecy rate.  investigates a three node network including only one user. Paper  considers a multiuser MISO downlink system with the presence of passive eavesdroppers and potential eavesdroppers. The problem of minimizing the total transmit power takes into account artificial noise and energy signal generation for protecting the transmitted information against both considered types of eavesdroppers. The semi-definite programming (SDP) relaxation approach is adopted to obtain the optimal solution. Both  and  use deterministic model for modelling the CSI uncertainty. Paper  proposes a linear precoder for a multiuser MIMO system in which multiusers potentially act as eavesdroppers. The proposed precoder is based on regularized channel inversion with a regularization parameter and power allocation vector. Then, an extension of the algorithm by jointly optimizing the regularization parameter and the power allocation vector is presented to maximize the secrecy sum-rate. Robust beamforming design in relay systems has also been investigated.  investigates the non-robust and robust cases of joint optimization in bidirectional multi-user multi-relay MIMO systems. The authors mainly concentrate on the sum MSE criterion as well as maximum user’s MSE. In , the relaying robust beamforming for device-to-device communication with channel uncertainty is considered.
In our paper, beamforming design of sum secrecy rate (SSR) optimization is investigated under sum power constraint for multiuser MISO channel. To solve the original complicated and nonconvex problem, an efficient approximation algorithm based on Taylor expansion is developed for the purpose of near-optimal solutions. In addition, a beamforming design scheme based on ZF with low complexity is presented. Numerical results demonstrate the better performance of the former algorithm. Both of the algorithms are compared to the SLNR algorithm, which mainly minimizes the power leaking to other user s channel space, and proved to be better.
The remaining part of this paper is organized as follows. The system model and the sum power constrained beamforming problem are presented in Section II. The solution based on Taylor expansion of the approximation problem under the assumption of imperfect CSI is discussed in detail in Section III. In Section IV, the beamforming design scheme based on ZF at the BS is demonstrated. Simulation results are presented in Section V. Finally, the conclusions are drawn in Section VI.
Notation: In this paper, we use bold uppercase and lowercase letters to denote matrices and vectors, respectively. and denote the transpose and the conjugate transpose of a matrix or a vector, respectively. denotes the conjugate of a matrix or a vector which means . is the trace of a matrix. denotes the rank of a matrix. We use the expression if
is complex Gaussian distributed with mean
and variance. denotes the Frobenius norm. represents the property of semidefinite. . denotes the number of all combinations of different elements chosen from elements.
Ii System Model and Problem Formulation
Ii-a System Model
In this paper, we will investigate a network incorporating one BS and 2 users, or equivalently user-eaves pairs. In the user-eaves pair, one legitimate user is wiretapped by an eavesdropper as shown in Fig. 1. It is presumed that the BS is equipped with antennas where , and each user or eavesdropper with a single antenna. It is supposed that the BS serves the
wiretapped users while each eavesdropper attempts to wiretap the legitimate user in the same user-eaves pair. It is assumed that the BS only knows imperfect CSI of each user and eavesdropper due to limited feedback or other reasons, and the channel estimation error is norm-bounded. The BS employs transmit beamforming to communicate withusers. Let denote the information signal sent for the -th user at time t, and let be the corresponding beamforming vector. The received signal at the -th wiretapped user is given by
where denotes the channel vector between the BS and the -th user, and is additive Gaussian noise at the -th user satisfying . As seen from (1), each eavesdropped user suffers from the intracell interference in addition to the noise. It is presumed that all receivers employ single-user detection where the intracell interference is simply treated as background noise.
Under the assumption that each user is wiretapped by the specific eavesdropper, let the -th eavesdropper be the one wiretapping the -th user, which means the -th eavesdropper has the knowledge of . The received signal at the -th eavesdropper can be written as
where denotes the channel vector between the BS and the -th eavesdropper, and is additive Gaussian noise at the -th eavesdropper satisfying .
The instantaneous achievable rate of the -th user can be transformed into the following
and the rate achieved by the -th eavesdropper can be expressed as
Notice that in this paper, the channel is supposed to be imperfect. Under the assumption that the channel estimation error is norm-bounded, we have
where and are the estimated channel of users and eavesdroppers, while and are the channel estimation errors, respectively. and are the bounds of the norm of the channel estimation error. It is presumed that .
Based on the expressions above, the SSR is defined as
Notice that in this paper, sum power constraint should be satisfied which is
Ii-B Problem Formulation
The objective is to maximize the SSR with the constraint of sum transmitting power. Thus, the optimization problem can be formulated as
However, the BS only has the knowledge of the estimated channel and the bound of the norm of the channel estimation error, which would bring difficulties to the modeling of the SSR problem. To deal with this, the lower bound of the objective in (9) would be taken into consideration merely. Lemma 1 would be introduced first to get the lower bound of the objective.
: For the two problems in the following,
where and are given parameters, their solutions can be expressed as
Apparently, we have
where and are vectors. The Cauchy-Schwarz inequality states that for all vectors x and y of an inner product space, it is true that , where is the inner product. According to Cauchy-Schwarz inequality,
Then we have
The inequality holds with equality when and are linearly dependent.
It can be concluded that the upper bound of is at the point , and the lower bound of is at the point . ∎
Then, based on Lemma 1, the upper bound of is given by
Similarly, the lower bound of can be expressed as
Notice that in (II-B) and (II-B), the second order error terms is omitted since it is quite small compared with other terms. In (II-B), is usually small and is larger than 0. When is large enough, the lower bound, , may be smaller than 0, which has no practical meaning. Due to , the problem what is large enough is dependent on the channels and the beamformers. The sum of several terms makes it more difficult to analyses the effect of large . From the simulation results, we find that if , we may not be able to get the beamformers or the sum secrecy rate will be quite low. In this case, the channel estimation error is intolerant and our method may not be able to solve this problem.
Thus, the lower bound of the objective in (9) is as follows.
Problem (20) can be formulated as the following
In problem (9), the perfect CSI is not known to the BS. The BS only has the knowledge of the estimated channel and the bound of the norm of the channel estimation error. Under this condition, problem (9) can’t be solved since the parameters of the real channels in problem (9) are missing. Then, the problem to maximize the lower bound of SSR is investigated and problem (20) is formulated based on the estimated channel. If the BS can obtain perfect CSI, that means the bound of the norm of the channel estimation error is set to zero. Then problem (20) would be the same as problem (9).
The upper bound (17) and the lower bound (18) are important in the formulation of problem (20). It is not easy to estimate the tightness of the upper bound and the lower bound. In the simulation results, the solutions are obtained. Then we compute the lower bound of SSR through (20a) and the practical SSR through (9a). A comparison between them can illustrate that the gap between the lower bound of SSR and the practical SSR is small.
The traditional leakage-based beamforming scheme  might be applicable to such situation, which mainly minimizes the power leaking to other user s channel space. The solution is given by , where . The norm of is adjusted according to the transmit power. Besides the signal-to-leakage-and-noise ratio (SLNR) scheme, our proposed algorithms are illustrated in Section III and Section IV. The comparisons between the proposed algorithms and SLNR algorithm shown in section V will demonstrate that the proposed algorithms are better.
Iii Approximation Method Based on Taylor Expansion
In this section, an efficient approximation algorithm is developed for the purpose of local-optimal solutions of problem (20). Since the objective function (20a) is nonconvex and complicated, problem (20) is difficult to solve. A convex approximation method based on Taylor expansion will be presented to handle problem (20) efficiently in the following.
Iii-a Convex Approximation
To make the problem more tractable, new matrixs are introduced. Thus, we have . It should be noticed that if instead of is used as the variables to optimize, has to satisfy , which would violate the problem’s convexity. Then, SDR (semidefinite relaximation), a convex optimization based approximation technique, is applied to omit the rank-one constraint. The approximated problem can be transformed into the following
It should be mentioned that SDR has been widely used in various beamforming design problems. If the solution of the problem satisfies the rank-one constraints, then eigenvalue decomposition would be utilized to obtain the practical optimal solution; otherwise, randomization technique can be applied .
for . Note that , which means the lower bound and upper bound of should be no less than 0. Then we have
Due to the sum power constraint, , , and won’t be infinity. It can be observed that , , and are bounded.
The objective function can be transformed into . It can be seen that the objective function is convex. Notice that the equalities in (22) have been replaced by inequalities as in (24b) to (24e). It could be verified by the monotonicity of the objective function that all the inequalities in (24b) to (24e) hold with equalities at the optimal points. To be specific, in the process of solving problem (24), if the inequalities (24b) or (24e) doesn’t hold with equality, we can increase or until the equality holds. If the inequalities (24c) or (24d) doesn’t hold with equality, we can decrease or until the equality holds. In the mean time, the optimal value of problem (24) would also be improved. Thus, the inequalities from (24b) to (24e) would hold with equalities for the final solutions.
For the purpose of maximization of (24a), we maximize and which are the lower bound of and , respectively, while minimizing and which are the upper bound of and , as can be observed from (24b) to (24e). Thus, while solving problem (24), the lower bound of the numerator of the objective function (21a) are maximized and the upper bound of the denominator are minimized leading to the maximization of (21). As explained above, it can be seen that problem (24) is an appropriate approximation of problem (III-A).
for . Then and are feasible to problem (24). Aiming to make (24c) and (24d) convex, these constraints are conservatively approximated at the point based on Taylor expansion . The Taylor series of a function that is infinitely differentiable at a number is the power series . Since both of and are convex, their first-order Taylor expansion at and are respectively given by
Through the first-order Taylor expansion of and , the original non-linear terms successfully turn out to be linear leading to convex constraints. By replacing (24c) and (24d) with (28a) and (28b), respectively, the following approximation of problem (24) can be obtained
Problem (29) is a convex optimization problem which can be efficiently solved by CVX, a package for specifying and solving convex programs ,.
In summary, the reformulation above consists of two approximation steps: a) the rank relaxation of to through SDR, and b) constraint restrictions of (24b) and (24e) to (29b) and (29e). Note that if problem (29) yields a rank-one optimal , a rank-one beamforming solution can be readily obtained by rank-one decomposition of for . It is then straightforward that this rank-one beamforming solution is also feasible to the original problem (20). Otherwise, randomization technique can be applied.
Iii-B Successive Convex Approximation
Formulation (29) is obtained by approximating problem (20) at the given feasible point , as described in (28). This approximation can be further improved by iterative procedure based on the optimal solution obtained through solving (29) in the previous approximation. Specifically, in the -th iteration, the following convex optimization problem is solved by CVX,