A Deep-learning-based Joint Inference for Secure Spatial Modulation Receiver
Feng Shu, Lin Liu, Yumeng Zhang, Guiyang Xia, Xiaoyu Liu, Jun Li, Shi Jin, Jiangzhou Wang
aa r X i v : . [ ee ss . SP ] J u l A Deep-learning-based Joint Inference for SecureSpatial Modulation Receiver
Feng Shu, Lin Liu, Yumeng Zhang, Guiyang Xia,Xiaoyu Liu, Jun Li, Shi Jin, and Jiangzhou Wang.
Abstract —As a green and secure wireless transmission way,secure spatial modulation (SM) is becoming a hot research area.Its basic idea is to exploit both the index of activated transmitantenna and amplitude phase modulation (APM) signal to carrymessages, improve security, and save energy. In this paper, wereviewed its crucial techniques: transmit antenna selection (TAS),artificial noise (AN) projection, power allocation (PA), and jointdetection at desired receiver. To achieve the optimal performanceof maximum likelihood (ML) detector, a deep-neural-network(DNN) joint detector is proposed to jointly infer the index oftransmit antenna and signal constellation point with a lower-complexity. Here, each layer of DNN is redesigned to optimize thejoint inference performance of two distinct types of information:transmit antenna index and signal constellation point. Simulationresults show that the proposed DNN method performs 3dB betterthan the conventional DNN structure and is close to ML detectionin the low and medium signal-to-noise ratio regions in terms ofthe bit error rate (BER) performance, but its complexity is farlower-complexity compared to ML. Finally, three key techniquesTAS, PA, and AN projection at transmitter can be combined tomake SM a true secure modulation.
I. S
ECURE SPATIAL MODULATION AND DEEP LEARNING
Spatial modulation (SM) concept was first proposed byChau and Yu in [1]. They had creatively proposed the conceptof SM: carry additive bit information via antenna indices. In[2], the authors made a systematic and in-depth investigationof SM, and officially named it as SM. At the same time, thebasic principle of SM was also explained. SM exploits boththe index of activated transmit antenna and amplitude phasemodulation (APM) signal to carry messages. Compared toBell Laboratories Layer Space-Time (BLAST) and space timecoding (STC), SM system achieves a good balance betweenspatial multiplexing and diversity. We call it as the third waybetween BLAST and STC. Compared to BLAST and STC,SM has a good advantage of high energy efficiency (EE) dueto the use of less active RF chains. Thus, it is a green wirelesstransmission technique.Wireless communication is usually prone to passive eaves-dropping and active malicious attacks due to its broadcastcharacteristics. Although there is a series of mature encryptionalgorithms in the upper layer of network protocol, it is stillfragile in wireless communication. To address this issue, thephysical layer security (PLS) technology becomes a naturechoice, and enhances its security from the perspective ofinformation theory. PLS has been extensively studied in [3].PLS will work with traditional cryptography to play a key roleand provide an incremental guarantee for the future personalprivacy protection and information network security. Working with encryption together, a dual protection of transmittingconfidential messages (CMs) will be achieved.In the past decade, secure modulation emerges as an specialform of multiple-input-multiple-output (MIMO). It mainlyconsists of two types: directional modulation (DM) and secureSM(SSM). DM using beamforming with the help of artificialnoise (AN) can securely deliver confidential messages (CMs)to desired users in line-of-sight channel, and is unsuitable forfading channels. Instead, SM is naturally suitable for fadingchannel.By introducing security into the SM, it is able to transmitCMs over the fading channel. However, transmitting CMs viaSM is an attractive and very important issue [4] [5] [6]. In [5],the authors made an extensive investigation of TAS methodsin SSM systems. Then, two high-performance transmissionantenna selection schemes: leakage-based and maximum SR,have been proposed to improve the SR performance, and thegeneralized Euclidean distance-optimized antenna selectionmethod has been generalized to provide a secure transmission.In [6], an active antenna-group (AAG) selection is proposed tomaximize the average secrecy rate (SR) in the case of limitedactive antenna pattern and finite-alphabet inputs.In SSM, how to optimize and design the AN projectionmatrix has a substantial impact on the SR performance. In[6], [7], AN was projected onto the null-space of the desiredchannel to improve the security of communication systems.The main benefit of this scheme is that the AN projectionmatrix has a closed-form expression. However, such a schememight cause some secrecy performance loss due to lack a holis-tic consideration of secure communication systems. In otherwords, the design of the beamformer of AN in conventionalway is merely achieved from the perspective of the desiredreceiver.Intelligent communication is considered as one of themainstream directions of the follow-up development of mo-bile communication after 5G. The basic idea is to introduceintelligent elements into all layers of mobile communicationsystem, realize the organic integration of mobile communica-tion and artificial intelligence technology, and greatly improvethe efficiency of mobile communication system. The previousresearch results have focused on the application layer andthe network layer, and the main idea is to introduce machinelearning, especially deep learning (DL), into wireless resourcemanagement, channel decoding, and other fields. DL is oneof the most important breakthroughs in the field of artificialintelligence in the past decade. The author details the DLalgorithms in [8]. It has been successfully applied in many
Fig. 1. Schematic diagram of secure SM network. fields such as computer vision, natural language processing,speech recognition, etc., and has achieved great success. Dueto the new features of future communication, such as complexscenes with unknown channel models, high-speed and accurateprocessing requirements, many scholars have introduced DLinto the physical layer of wireless communication [9]. In thephysical layer, there is a new trend of combining wirelesstransmission and DL.In [10], the authors considered channel estimation formillimeter-wave massive MIMO systems. An approximatemessaging network based on learning denoising was proposedfor channel estimation, which can learn channel structure andestimate channel from a large amount of training data. In[11] a new framework were proposed for integrating large-scale MIMO and DL to address the problem of channelestimation and DOA estimation [12]. Deep neural network(DNN) is used for offline learning and online learning, and thestatistical characteristics of the wireless channel and the spatialcharacteristics of the angle domain are effectively learned.A novel deep learning assisted sparse coded multiple accessscheme is proposed in [13]. By using a DNN-based encoderand decoder adaptively construct a codebook that minimizesthe bit error rate in [14].In Fig. 1, a typical secure SM system is shown. In thisfigure, four main tools including transmit antenna selection(TAS), beamforming of confidential messages, AN projection,and power allocation (PA) are fully utilized to achieve a SSM.In such a network, at desired receiver, the joint detectionof transmit antenna index and signal constellation point isrequired. Joint detection performance is very important. Con-ventionally, the optimal maximum likelihood (ML) detectoris a natural choice. But, as the number of transmit antennastends to medium-scale or large-scale or the size of signalconstellation goes to medium-scale or large-scale, the MLjoint detector is confronted with a complexity bottleneck, i.e.,exponential complexity. To reduce the computational complex-ity of receiver, the DNN-based joint detector is proposed tojointly infer the transmit index and signal constellation pointin this paper. Compared to ML, the joint DL detector is low- complexity and reach the optimal performance of ML.II. S
YSTEM MODEL AND T RANSMIT ANTENNA SELECTION
Consider a typical SSM system as shown in Fig. 1. Inthis system, there is a transmitter (Alice) equipped with N a transmit antennas. Without loss of generality, when thenumber of antennas at the transmitter is not a power of two, N t = 2 ⌊ log Na ⌋ , out of N a transmit antennas are selectedfor mapping the bits to the antenna index. The log M bitsare used to form a constellation symbol, where M is thesignal constellation size. As a result, the spectral efficiencyis log N t + log M bits per channel use (bpcu).Referring to the secure SM model in [15], the transmitsignal vector with the aid of AN can be given by x = p βP S e n s m + p (1 − β ) P S P AN n (1)where P S denotes the total transmit power constraint and β isthe PA factor. e n is the n -th column of identity matrix I Nt ,and s m is the digital constellation symbol with a normalizedpower E [ | s m | ] = 1 . Additionally, P AN is the AN projectionmatrix and n ∈ C N t × is the corresponding AN vector. Then,the receive signal at desired receiver Bob can be formulatedas follows y d = p βP S HT k e n s m + p (1 − β ) P S HT k P AN n + n b (2)where H are the complex channel gain matrix from Alice toBob and to Eve, T k is transmit antenna selection matrix.Selecting an active antenna group can be adopted to furtherimprove the performance of SM systems. There are severalexisting TAS methods for secure SM system as follows: ran-dom, leakage [4], and generalized Euclidean distance antennaselection (EDAS). For the leakage-based TAS strategies, thesignal-to-leakage-and-noise ratio (SLNR) of CM from eachtransmit antenna is computed and formed a sequence ofSLNRs, where SLNR is defined as the ratio of the receivesignal power at Bob to the sum of the receive power ofCM at Eve, receive AN power, and channel noise variance.Then a low-complexity sorting algorithm places the valuesof SLNR in decreasing order. The antennas corresponding antennas associated with the top N SLNRs is chosen, calledMax-SLNR [4]. The Max-SLNR can achieve the near-optimalSR performance with a low-complexity.From the aspect of decoding performance at receiver, gen-eralized EDAS performs best in terms of bit error rate. Thegeneralized EDAS Method is aim to select a TAS patternof maximizing the minimum Euclidean distance over desiredchannel or minimizing the minimum Euclidean distance overeavesdropping channel due to the fact that the minimumdistance has a direct relationship to BER. -30 -20 -10 0 10 20 30 40
SNR (dB) S ec r ec y R a t e ( b it s / s / H z ) EDASLeakage method [4]Random methodES
Fig. 2. Comparison of SR performance of various TAS methods.
Fig. 2 demonstrates the SR performance comparison ofthe optimal exhaustive search (ES), Max-SLNR, EDAS, andrandom methods without the aid of AN. From this figure, itis seen that the four methods have a decreasing order in SRperformance as follows: ES, Max-SLNR, EDAS, and randommethod. Additionally, we also find an interesting result: allthe SR curves first go up as hills, then reach their peaks, andfinally go down hills as SNR increases. In other words, all theSR curves have main peaks, and can be approximately viewedas concave functions of the SNR.Fig. 3 makes a SR performance comparison of ES, Max-SLNR, EDAS, and random methods with the aid of AN.From this figure, it is seen that the four methods have stillan decreasing order in SR as follows: ES, Max-SLNR, EDAS,and random. In particular, observing Fig. 2 and Fig. 3, we findan important fact: with the aid of AN, the SR performance canbe improved significantly, especially in the medium and highSNR regions. The values of SR for the four methods growsgradually as SNR increases. When SNR enters the high SNRregion, their SR performance reaches their corresponding SRceils.III. B
EAMFORMING , AN
PROJECTION , AND P OWERALLOCATION
Since secure SM channel can be viewed as a discrete-inputcontinuous-output memoryless channels (DCMC), it is veryhard to find a closed-form expression for mutual informationin such a network. Mutual information contains the expected -10 -5 0 5 10 15 20
SNR (dB) A v e r a g e S ec r ec y R a t e ( b it s / s / H z ) Random methodEDASESLeakage method [4]
Fig. 3. Comparison of SR performance of various PA strategies. items of noise, usually with high computational complexity.Only in the case that the input is a Gaussian signal, The SRexpression will have a very concise form.For DCMC, in general, an approximate estimation SRexpression is used instead of exact SR, it is difficult to convertto a convex problem. There is an inversion in the expression,and the outer layer needs to solve the expectation. This isa rather complicated issue. In addition, it is difficult for thetransmitter to optimize the CM beamforming vector and ANprojection matrix by maximizing SR. This is a a challengingproblem in the coming future. However, at the cost of someSR performance loss, some low-complexity and closed-formmethods can be used. For example, the AN projection matrixcan be constructed from the null-space of the desired channelfrom Alice to Bob while the CM beamforming vector is alsoformed from the null-space of the desired channel from Aliceto Eve. If you want to further the SR performance, the leakage-based rule is used to optimize the design of AN projectionmatrix and CM beamforming vector.PA, as an efficient way to enhance security, has beeninvestigated in [15]. By adjusting the PA factor, the powercan be allocated between CM and AN freely to affect the SRand BER performance. By simulation and proof, it confirmsthat SR is a concave function of the PA factor β . Althoughexhaustive search (ES) can be employed to search the optimalPA factor, its computational complexity is high. Thus, in [15],a novel PA strategy, called Max-P-SINR-ANSNR where ’P’is short for product, and ’ANSNR’ stands for AN-to-signal-plus-noise ratio, presented a closed-form expression for thePA factor. This dramatically reduces the complexity of ES.Fig. 4 makes a comparison of several typical PA strategies:ES, fixed, gradient descent (GD), and Max-P-SINR-ANSNR.From this figure, it follows that the Max-P-SINR-ANSNR andGB strategies in [15] are close to the optimal ES, but theircomplexity is dramatically lower than ES. Comparing the threemethods with fixed power allocation factors, it can be seen thatthe SR at β = 0 . is the lowest one, and β = 0 . is the highestone in the value of high SNR. This is because when the SNR is SNR(dB) A v e r a g e S ec r ec y R a t e ( b it s / s / H z ) Max-P-SINR-ANSNR [15]GD=0.50=0.25=0.1ES
Fig. 4. Comparison of SR of various PA strategies. high, both Bob and Eve have a very good quality of channel,and a large portion of transmit power may be allocated toAN to disturb eavesdropper, so as to obtain a high securityperformance. Additionally, due to its closed-form expressionof Max-P-SINR-ANSNR, it strikes a good balance betweencomplexity and performance.IV. P
ROPOSED
DNN-
BASED J OINT I NFERENCE OFANTENNA INDEX AND CONSTELLATION POINT
Assuming Bob has the perfect channel state information(CSI) of H and T k , the joint ML detector (MLD) at desiredreceiver can be casted as follows [ˆ n, ˆ m ] = arg min n ∈ [1 ,N t ] ,m ∈ [1 ,M ] (cid:13)(cid:13)(cid:13) y d − p βP S HT k e n s m (cid:13)(cid:13)(cid:13) (3)where ˆ n and ˆ m denote the index of transmit antenna andsignal constellation point by joint MLD, respectively. TheMLD requires the computational complexity N T AS × N SC floating-point operations (FLOPs) where N T AS and N SC denote the complexity of TAS scheme and the constellationsize, respectively. Obviously, this complexity is a product. As N T AS , N SC or both tends to large-scale, the total complexitywill become a large number.To reduce this complexity, a DNN is a natural choice due tothe fact that determining which antenna or constellation pointis actually a kind of classifying. However, the conventionalDNN structure shown in Fig. 5, locating on the right-uppercorner, is verified to be 3dB worse than the ML in termsof BER given a fix BER= − . To completely remove the3dB performance gap, a novel DNN structure shown in theleft-bottom corner of Fig. 5 is proposed. Here, each layer isredesigned to have three kinds of outputs: S k , E k and V k ,where S k and E k are the estimate of constellation symbol andantenna index in the k -layer, respectively. V k is the hiddenoutput vector of the k -layer, and is also the input of the nextlayer. The DNN-based joint inference idea is to minimize theloss function k x − ˆ x ( y d , H ) k = k x − F ( y d , H ; w , b ) k (4) by random GD and back-propagation methods, where F ( y d , H ; W , b ) = f n − (cid:18) W n − f n − (cid:18) · · · (cid:18) W f (cid:18) W (cid:18) y d Vec ( H ) (cid:19) + b (cid:19) + · · · (cid:19) + b n − (cid:19) (5), where W k is the matrix of weight coefficients correspondingto layer k , and b k stands for the bias vector for the correspond-ing layer. f ( k ) ( · ) represents nonlinear activation function oflayer k and describes an input-output mapping. Vec ( H ) isthe vectorization of the matrix H To solve the gradient disappearance problem, some newactivation functions are adopted to replace the classical sig-moid activation function. Among them, the Rectified LinearUnit (ReLU) is commonly used. Introducing nonlinearity byzeroing a function value less than zero. When the input valueis greater than zero, the function is a linear function, whichsimplifies the gradient calculation and the gradient value willnot decrease with the increase of the input. This alleviates thegradient disappearance problem.To speed up the convergence and reduce the computationalcomplexity, the classical DG algorithm adjusts to the fall ofthe stochastic gradient. that is, randomly selects a sample tocalculate the loss function and the gradient. However, therandom selection of sample selection causes a large volatilityin the training process, that is, the network cannot convergeto the optimal solution, but fluctuates around the optimalsolution. Take a compromise, using a batch gradient descentmethod, each time a batch of samples is selected to calculatethe loss function and gradient. It’s also not guaranteed tobe globally optimal by this method, but we can receive thissolution as the loss function is reduced to be small.Even using the above two strategies, neural networks are notvery easy to train. Then, some of the current popular neuralnetwork training techniques can be adopted to improve itsperformance, speed and stability. The use of a moving averagemodel enhances the stability of the parameters. Dropping outunits (both hidden and visible) by a probability when trainingneural networks can efficiently address the problem of networkover-fitting and improve its generalization ability. Additionally,the batch normalization is applied to improve the performance.Now, we present numerical results to validate the effec-tiveness of our proposed DNN-based scheme. Constructingtraining and validation sets: generating a large amount of train-ing data by performing baseband numerical simulation on theSM-MIMO channel model, and performing two-dimensionallabeling of the antenna and the constellation point on thereceived vector data. The structure of the neural networkdepth and the number of neurons per layer are determined byinitializing the weight matrix and bias of the network. Thenhyperparameters: maximum number of training, batch size,initial learning rate, and drop probability, are also initialized.The network is trained by training set, and the verificationset is used to verify the network after every 100 iterations todecide whether to terminate early. Apply the above learningcompletion network to a desired receiver of SM to verify itsBER performance. (cid:44)(cid:81)(cid:83)(cid:88)(cid:87)(cid:3)(cid:79)(cid:68)(cid:92)(cid:72)(cid:85) (cid:43)(cid:76)(cid:71)(cid:71)(cid:72)(cid:81)(cid:3)(cid:79)(cid:68)(cid:92)(cid:72)(cid:85)(cid:86) (cid:50)(cid:88)(cid:87)(cid:83)(cid:88)(cid:87)(cid:3)(cid:79)(cid:68)(cid:92)(cid:72)(cid:85) (cid:266)(cid:266)(cid:266)(cid:266)(cid:266) (cid:47)(cid:68)(cid:92)(cid:72)(cid:85)(cid:3)(cid:20)(cid:47)(cid:68)(cid:92)(cid:72)(cid:85)(cid:3)(cid:78)(cid:47)(cid:68)(cid:92)(cid:72)(cid:85)(cid:3)(cid:81) (cid:258) (cid:54) (cid:78) (cid:40) (cid:78) (cid:57) (cid:78) (cid:92)(cid:40) (cid:78)(cid:16)(cid:20) (cid:43)(cid:54) (cid:78)(cid:16)(cid:20) (cid:57) (cid:78)(cid:16)(cid:20) N b (cid:20) Fig. 5. Proposed novel DNN structure for joint inference of antenna index and constellation point.
SNR(dB) -6 -5 -4 -3 -2 -1 B E R MMSEZFConventional DNNProposed novel DNNML
Fig. 6. BER performance comparison among proposed DNN, joint ML,conventional DNN, ZF, and MMSE with N a =4, N r =4 and QPSK. Fig. 6 illustrates the curves of average BER versus SNRfor the five methods with N t = 4 , N r = 4 , and modulationbeing QPSK, where several classic detector such as ML, zero-forcing, minimum mean square error (MMSE) are used as per-formance benchmarks. From Fig. 6, it is seen that the proposednovel DNN-based method approaches the ML method whenSNR is lower than 10dB, and much better than ZF, MMSE andconventional DNN ones in almost all SNR regions in terms ofBER performance. It is evident that the conventional DNN-based method is 3dB worse than the proposed novel DNNand ML ones at BER = 10 − . In summary, the network’sperformance can be significantly improved by adjusting thestructure of DNN. V. O PEN PROBLEMS
There are still many open problems existed to be addressed.Here, we list several important ones of them as follows:1) As the number of antennas at SM transmitter tends tolarge-scale, the circuit cost and complexity becomes asignificant obstacle for practical applications of SM. Ahybrid analog-digital structure is preferred. In such ascenario, how to achieve an optimal strategy of transmitantenna subarray is challegening problem. This problemcan be modelled an integer optimization problem. Thekey is to develop a low-complexity algorithm.2) If Eve works in a full-duplex model and becomes anactive eavesdropper, i.e., jamming, how to optimize thedesign of the transmitter at Alice in order to reduce theeffect of jamming from Eve and at the same time achievea feasible performance is a hard task. For Bob, thereceive beamforming scheme of combating the jammingis preferred.3) By adjusting the neural network structure of deep learn-ing, the BER performance of the proposed structure withQPSK is close to the performance of ML detection. Toextend it to the high-order modulation is still an openproblem. This requires us to optimize the structure ofdeep learning.4) In the presence of CSI measurement errors, how toconstruct robust beamforming, PA, and TAS by takingthe statistical property of CSI error into account requiresa great effort. In particular, the first task is to formulatethe SR expression or approximate expression in such ascenario. This will pave a way for robust beamforming,PA, and TAS.5) Due to non-convexity of the SR expression, AN opti-mization is a challenging task because the beamformerof AN is always included in a complicated expression, especially when a rough/statistic CSI of Eve’s channelis available at the transmitter.6) As the number of TAs further increases, how to optimizean AAG for enhancing the security of the communi-cation systems becomes intractable, this is because thecomputational complexity of evaluating the accurate SRwill be explosively grown upon increasing the numberof TAs. That requires us exploring some concise metricsinstead of the original function to evaluate the value ofSR. VI. C
ONCLUSION
In this article, the great potential of secure SM has beenhighlighted as a key secure tool for future vehicular commu-nications, IoT, UAV, smart transportation, and satellite com-munications. We review its key techniques: TAS schemes, PAstrategies, and joint detection methods at desired receiver. Anew DNN structure was proposed to jointly infer the transmitantenna index and signal constellation point, i.e., a new jointdetection way with low-complexity. Also, we have raisedseveral new open important future research problems. Finally,in our view, secure SM will have wide diverse promisingapplications in the near future.R
EFERENCES[1] Y. A. Chau and Shi-Hong Yu, “Space modulation on wireless fadingchannels,” in
IEEE 54th Vehicular Technology Conference. VTC Fall2001. Proceedings (Cat. No.01CH37211) , vol. 3, Oct 2001, pp. 1668–1671 vol.3.[2] R. Y. Mesleh, H. Haas, S. Sinanovic, C. W. Ahn, and S. Yun, “Spatialmodulation,”
IEEE IEEE Trans. Veh. Technol. , vol. 57, no. 4, pp. 2228–2241, July. 2008.[3] W. Trappe, “The challenges facing physical layer security,”
IEEE Com-mun. Mag. , vol. 53, no. 6, pp. 16–20, June 2015.[4] F. Shu, Z. Wang, R. Chen, Y. Wu, and J. Wang, “Two high-performanceschemes of transmit antenna selection for secure spatial modulation,”
IEEE Trans. Veh. Technol. , vol. 67, no. 9, pp. 8969–8973, Sep. 2018.[5] F. Shu, X. Wu, J. Hu, J. Li, R. Chen, and J. Wang, “Secure and precisewireless transmission for random-subcarrier-selection-based directionalmodulation transmit antenna array,”
IEEE J. Sel. Areas Commun. ,vol. 36, no. 4, pp. 890–904, Apr. 2018.[6] G. Xia, F. Shu, Y. Zhang, J. Wang, S. ten Brink, and J. Speidel, “Antennaselection method of maximizing secrecy rate for green secure spatialmodulation,”
IEEE Trans. Green Commun and Netw. , vol. 3, no. 2, pp.288–301, Jun. 2019.[7] F. Wu, L. Yang, W. Wang, and Z. Kong, “Secret precoding-aided spatialmodulation,”
IEEE Commun. Lett. , vol. 19, no. 9, pp. 1544–1547, Sep.2015.[8] J. Schmidhuber, “Deep learning in neural networks: An overview,”
Neural Netw , vol. 61, pp. 85–117, 2015.[9] C. Jiang, H. Zhang, Y. Ren, Z. Han, K. Chen, and L. Hanzo, “Machinelearning paradigms for next-generation wireless networks,”
WirelessCommun. , vol. 24, no. 2, pp. 98–105, Apr. 2017.[10] H. He, C. Wen, S. Jin, and G. Y. Li, “Deep learning-based channelestimation for beamspace mmwave massive mimo systems,”
IEEEWireless Communications Letters , vol. 7, no. 5, pp. 852–855, Oct. 2018.[11] E. Nachmani, Y. Be’ery, and D. Burshtein, “Learning to decode linearcodes using deep learning,” in , Sep. 2016, pp.341–346.[12] H. Huang, J. Yang, H. Huang, Y. Song, and G. Gui, “Deep learning forsuper-resolution channel estimation and DOA estimation based massiveMIMO system,”
IEEE Trans. Veh. Technol. , vol. 67, no. 9, pp. 8549–8560, Aug. 2018.[13] M. Kim, N. Kim, W. Lee, and D. Cho, “Deep learning-aided scma,”
IEEE Commun. Lett. , vol. 22, no. 4, pp. 720–723, Apr. 2018. [14] E. Nachmani, E. Marciano, L. Lugosch, W. J. Gross, D. Burshtein,and Y. Beery, “Deep learning methods for improved decoding of linearcodes,”
IEEE J. Sel. Topics Signal Process. , vol. 12, no. 1, pp. 119–131,Feb. 2018.[15] F. Shu, X. Liu, G. Xia, T. Xu, J. Li, and J. Wang, “High-performancepower allocation strategies for secure spatial modulation,”
IEEE Trans.Veh. Technol. , vol. 68, no. 5, pp. 5164–5168, May 2019.FENG SHU is a professor with the School of Electronic and OpticalEngineering at Nanjing University of Science and Technology, Nanjing, China.He is also with the College of Physics and Information, Fuzhou University,Fuzhou 350116, China, and the College of Computer and information atFujian Agriculture and Forestry University, Fuzhou 350002, China. He hasbeen awarded with Mingjian Scholar Chair Professor and Fujian Hundred-talents Program in Fujian Province, China. He has published more 200 journalpapers on signal processing and communications, with more than 100 SCI-indexed papers and more than 80 IEEE journal papers. Her research interestsinclude MIMO and beamforming technologies, machine learning for mobilecommunications and multiple access techniques.LIN LIU is a postgraduate student in the School of Electronic and OpticalEngineering at Nanjing University of Science and Technology, Nanjing, China.His research interests include massive MIMO, wireless localization, and DOAmeasurement in wireless communications.YUMENG ZHANG is a graduate student in the School of Electronic and Op-tical Engineering at Nanjing University of Science and Technology, Nanjing,China. Her research interests include massive MIMO, wireless localization,and DOA measurement in wireless communications.GuiYANG Xia is a PhD student in the School of Electronic and OpticalEngineering at Nanjing University of Science and Technology, Nanjing, China.His research interests include massive MIMO, spatial modulation, and DOAmeasurement in wireless communications.XIAOYU LIU is a postgraduate student in the School of Electronic and OpticalEngineering at Nanjing University of Science and Technology, Nanjing, China.Her research interests include massive MIMO, wireless localization, and DOAmeasurement in wireless communications.JUN LI is a professor with the School of Electronic and Optical Engineering atNanjing University of Science and Technology, Nanjing, China. His researchinterests include caching, computing, and channel coding., vol. 68, no. 5, pp. 5164–5168, May 2019.FENG SHU is a professor with the School of Electronic and OpticalEngineering at Nanjing University of Science and Technology, Nanjing, China.He is also with the College of Physics and Information, Fuzhou University,Fuzhou 350116, China, and the College of Computer and information atFujian Agriculture and Forestry University, Fuzhou 350002, China. He hasbeen awarded with Mingjian Scholar Chair Professor and Fujian Hundred-talents Program in Fujian Province, China. He has published more 200 journalpapers on signal processing and communications, with more than 100 SCI-indexed papers and more than 80 IEEE journal papers. Her research interestsinclude MIMO and beamforming technologies, machine learning for mobilecommunications and multiple access techniques.LIN LIU is a postgraduate student in the School of Electronic and OpticalEngineering at Nanjing University of Science and Technology, Nanjing, China.His research interests include massive MIMO, wireless localization, and DOAmeasurement in wireless communications.YUMENG ZHANG is a graduate student in the School of Electronic and Op-tical Engineering at Nanjing University of Science and Technology, Nanjing,China. Her research interests include massive MIMO, wireless localization,and DOA measurement in wireless communications.GuiYANG Xia is a PhD student in the School of Electronic and OpticalEngineering at Nanjing University of Science and Technology, Nanjing, China.His research interests include massive MIMO, spatial modulation, and DOAmeasurement in wireless communications.XIAOYU LIU is a postgraduate student in the School of Electronic and OpticalEngineering at Nanjing University of Science and Technology, Nanjing, China.Her research interests include massive MIMO, wireless localization, and DOAmeasurement in wireless communications.JUN LI is a professor with the School of Electronic and Optical Engineering atNanjing University of Science and Technology, Nanjing, China. His researchinterests include caching, computing, and channel coding.