With the future substantial increase in coverage and network heterogeneity, emerging networks will encounter unprecedented security threats. Covert communication is considered a potential enhanced security and privacy solution for safeguarding future wireless networks, as it can enable monitors to detect the transmitter’s transmission behavior with a low probability, thereby ensuring the secure transmission of private information. Due to its favorable security, it is foreseeable that covert communication will be widely used in various wireless communication settings such as medical, financial, and military scenarios. However, existing covert communication methods still present many challenges toward practical applications. In particular, it is difficult to guarantee the effectiveness of covert schemes based on the randomness of eavesdropping environments, and it is challenging for legitimate users to detect weak covert signals. Considering that emerging artificial-intelligence-aided transmission technologies can open up entirely new opportunities to address the above challenges, we provide a comprehensive review of recent advances and potential research directions in the field of intelligent covert communications in this work. First, the basic concepts and performance metrics of covert communications are introduced. Then, existing effective covert communication techniques in the time, frequency, spatial, power, and modulation domains are reviewed. Finally, this paper discusses potential implementations and challenges for intelligent covert communications in future networks.
With the approach of the big data era, massive amounts of private information will be transmitted through wireless communication systems. However, the transmission of vital information presents severe security threats due to the openness and broadcasting properties of wireless channels. Especially in specific communication scenarios with a high security level, such as government affairs and military operations, the mere exposure of communication behavior may bring unpredictable risks and losses. Covert communication [1]—also known as low probability of detection (LPD) communication [2]—is a technique based on information theory that has attracted widespread attention for its particular ability to conceal communication behavior. By modeling samples of the monitor’s observed signals as a binary hypothesis testing problem, analyzing the monitor’s detection performance, and constructing a covert transmission scheme, covert communication decreases the probability of the monitor detecting the transmitter’s transmission behavior, thus ensuring the secure transmission of private information [1].
Before the development of covert communication techniques, the main schemes for ensuring secure data transmission were cryptography-based encryption at the upper layer [3] and physical layer security (PLS), which exploited the differences between legitimate and eavesdropping channels, as well as the randomness of the wireless channel [4]. Unlike traditional encryption techniques, covert communication is not limited by the robust information processing and computational capabilities (e.g., quantum computing) of the eavesdropping side and avoids the extra overhead required for distributing and managing keys. In addition, compared with PLS techniques, which only protect the content of confidential communications, covert communication can safeguard a more basic and broader range of legitimate communication behaviors from being interfered with or deciphered.
In recent years, fundamental theoretical studies have been continuously developed and enhanced. In particular, the square root law reveals zero covert capacity while setting theoretical bounds on covert transmission. As a result, a theoretically solid and widely applicable analytical framework for covert communication has been established [1], greatly facilitating research on the performance analysis and implementation optimization of techniques for achieving positive covert capacity in different scenarios. In the traditional covert communication research and application process, the secure transmission of data is mainly achieved through various types of covert means based on the time, frequency, space, power, and modulation domains. Among these, a random transmission time scheme has been used to confuse the binary detection of covert communications [5], [6], [7], [8]. Direct-sequence spread spectrum and frequency-hopping (FH) schemes have been shown to have low detection probability [9], [10], [11], [12], [13]. Multiple antenna techniques, terahertz schemes, and frequency diversity array techniques based in the spatial domain have also been widely used to enhance covert communications [14], [15], [16], [17], [18], [19]. Covert communication can also be achieved by increasing the uncertainty through power domain strategies such as power adaption and artificial noise (AN) [20], [21], [22], [23], [24], [25]. The covert capacity can be further enhanced by covert communication schemes based on random digital modulation and waveform-superposition techniques [26], [27], [28], [29]. Nowadays, typical applications of covert communication include cognitive radio networks, Internet of Things (IoT) networks, intelligent reflecting surface (IRS)-assisted networks [30], unmanned aerial vehicle (UAV)-assisted networks [31], and low-Earth orbit (LEO) satellite networks.
However, existing covert communication approaches still have obvious deficiencies in application; for example, current research generally considers an ideal situation in which the eavesdropping side knows the information of the legitimate user in advance, while the legitimate user also knows the location of the eavesdropper and the detection information in advance. However, owing to the uncertainty of the eavesdropper’s location, the unavailability of channel state information (CSI), the requirement of real-time detection, and so forth, the covert performance that can be truly obtained by the legitimate user only relies on the existing covert probability metrics, which cannot be completely and accurately metricized. In particular, both legitimate users and eavesdroppers urgently need to make real-time adjustments to their scenarios according to each other’s situation, making it vital to explore a new paradigm for covert communications.
Intelligent covert communication—that is, the use of intelligent technological methods to achieve more efficient and secure covert transmission—has become a very promising research direction and application. For example, with significant advances in machine learning (ML) techniques, it has become feasible to determine CSI from the environmental information of a wireless link [26]. In particular, mapping functions can well mimic some of the models employed in ML. These models can be accurately trained with a large number of real data samples to extract useful environmental information in the absence of explicit feedback/detection, which can be used to design and optimize covert transmission strategies to assist in the secure transmission of covert information [32]. ML algorithms can also be combined to achieve real-time adjustment of the design of an IRS, dynamically adjusting the phase, amplitude, and other characteristics of the reflected signals to achieve directional propagation or attenuation of the signals in order to improve the covertness of the communication link [33]. Signal propagation or attenuation can be realized in a directional way to improve the concealment of the communication link [34]. In addition, new dimensional domain covert waveform-generation techniques have been developed based on orthogonal time–frequency space or affine frequency division multiplexing [35]. Intelligent cooperative covert communication (ICCC) schemes [36], intelligent parasitic covert transmission schemes [37], and integrated sensing and communication (ISAC) covert transmission schemes [38] have also attracted widespread attention.
In view of the challenges and opportunities that arise in intelligent covert wireless communication, emerging wireless technologies—such as transform-domain signal processing, massive antennas, and ISAC—can be leveraged to achieve new levels of secrecy, privacy, and covertness. Motivated by these considerations, we provide a comprehensive review of recent advances and potential research in intelligent covert wireless communication in this survey. In particular, we concentrate on the following three aspects:
(1) The basics and fundamentals of covert communications. The concept of covert communications, the corresponding performance metrics, and the differences between it and other security techniques are introduced.
(2) Existing covert transmission strategies. These strategies are summarized in terms of the time, frequency, space, power, and modulation domains.
(3) Future research directions and challenges. Research directions and challenges for intelligent covert communication are suggested, including covert transmission enabled by artificial intelligence (AI) techniques, covert communication for integrated space–air–terrestrial networks, multidimensional covert waveform design, covert communication with active detection, and enhanced ISAC covert communication design.
2. The basics and fundamentals of covert communication
This section introduces basics and fundamental concepts of covert communications, including typical system models and common detection schemes.
2.1. The concept of covert communications
Covert communication is low-detection and low-interception communication that not only protects the concealed information but also makes it impossible for an adversary to detect the presence of the communication. More specifically, in covert communication, the signal modulation method and waveform characteristics are designed to hide the wireless transmission, in order to mitigate the threat of the user and the presence of the communication being detected [2].
In 2013, milestone research conducted by Bash et al. [1] introduced the concept of covert communication. For the first time, they derived the square root law of channel capacity under additive white Gaussian noise (AWGN), which is used as a fundamental information theoretical limit for LPD communication. The square root law states that no more than bits can be conveyed reliably from Alice to Bob, where denotes the number of channel uses to transmit secret message, while the lower-bounding error probability of Willie’s detection of this transmission is no less than a specific value .
Typical system model: Covert communication aims to enable a sender (Alice) to transmit data securely to an intended recipient (Bob), while ensuring that the act of communication remains undetectable by an adversary (Willie). For the adversary, assessing covertness involves quantifying how the statistical properties of the observed data differ in the presence or absence of communication. On the other hand, from Alice’s perspective, it is crucial to address the dual challenges of optimizing the communication performance of the main link (i.e., the Alice-to-Bob channel) and simultaneously ensuring sufficient levels of covertness. The signal observed by Willie in the th channel use () can be modeled under two hypotheses, as follows:
where and represent that Alice is keeping silent or transmitting, respectively. The secret message transmitted by Alice is represented as , while the environmental noise is modeled as . Typically, there are several ways for Willie to detect whether transmission behavior is occurring.
Willie determines whether Alice is transmitting () or remaining silent () by evaluating the average received signal power , which can be expressed as follows:
where denotes the preset power threshold at Willie. Willie faces two possibilities of error when analyzing the received signals: a false alarm (FA) and missed detection (MD). The probability of an FA is defined as , which can be described as Willie incorrectly concluding that Alice is transmitting when Alice is actually silent. Conversely, the probability of MD is given by indicating that Willie mistakenly believes that Alice is silent when she is actually transmitting. Willie’s primary objective is to minimize the total error probability by appropriately selecting his power detection threshold .
Aside from energy detection, sequential change-point detection (SCPD) is a commonly used detection scheme that can simultaneously collect observations and make judgments [39]. As shown in Fig. 1, by continuously analyzing the newly observed data, Willie determines whether there is a change point that differs from the previous data distribution. Once a change point is detected, it means that the system state has changed. In other words, the occurrence of communication between Alice and Bob introduces a statistical change in the sequence of signals received by Willie. For Willie, detecting whether communication is happening is essentially equivalent to identifying this change in statistical properties. Since Willie lacks prior knowledge about the exact timing of the communication, the detection process must operate sequentially. Therefore, at each time point, as a new sample is observed, Willie updates his decision based on all the samples collected up to that point. Given this framework, the design of Willie’s detector aligns with the principles of SCPD, also known as “quickest detection.” The detection process can be formalized in terms of the stopping time , which represents the time at which Willie decides to raise an alarm. The stopping time is defined as follows:
where denotes the lower bound, denotes the statistical value of the detection sample at time , and denotes the detection threshold. Similar to binary hypothesis testing, SCPD produces two outcomes: ① : Willie correctly detects Alice’s communication behavior; or ② : Willie issues an FA, where ( = 0, 1, 2, …) denote the time Alice starts to transmit. Therefore, the following two average runtimes are crucial for evaluating the performance of SCPD: the time interval during which Willie successfully detects the communication behavior, and the time interval during which Willie is alerted under the condition that no covert message is transmitted.
2.2. Covertness metrics
This section presents the covertness metrics. The performance of a covert communication system is measured from two perspectives: the error detection probability and detection probability from Willie’s perspective; and the covert outage probability and effective covert rate from the legitimate transceiver’s perspective. The design of covert transmission schemes is usually constrained by the error detection probability at Willie’s end, with the aim of achieving a higher effective covert rate.
2.2.1. Error detection probability
The error detection probability (Ped) for Willie’s optimal test is the sum of the probabilities of an FA () and MD (). This total error probability can be expressed as follows:
2.2.2. Detection probability
The detection probability represents the likelihood that Willie makes the correct decision. This occurs when Willie correctly decides (no transmission) under hypothesis (Alice is silent) and when Willie correctly decides (transmission) under hypothesis (Alice is transmitting). Mathematically, is defined as follows:
The error detection probability and detection probability can be used to measure not only the covert performance of the transmission between legitimate transceivers but also the performance of Willie in illegally detecting whether or not the legitimate users are transmitting covert information.
2.2.3. Covert outage probability
The covert outage probability is defined as the probability that the channel capacity cannot support a given target rate , which is given by
where is the bandwidth of the legitimate channel, and is the instant signal-to-noise ratio (SNR) of the legitimate channel.
From Eq. (7), when the system transmission rate is less than the target transmission rate, confidential information cannot be transmitted reliably and is considered to be an outage. The covert outage probability is usually used as a measure of the covert transmission by legitimate transceivers.
2.2.4. Effective covert rate
The effective covert rate (Rsb) at Bob’s end quantifies the average amount of information successfully transmitted while maintaining covertness. This can be mathematically expressed as follows:
The effective covert rate—that is, the achieved transmission rate without outage—is an important measure of the transmission capability of the legitimate users.
2.3. Differences from other security techniques
Table 1 summarizes the differences between covert communication and the aforementioned encryption, PLS, and spread spectrum communication techniques.
2.3.1. Encryption
Encryption is a classical and widely used technique for ensuring secure data transmission that operates at the upper layers of communication systems [3]. It works by applying encryption algorithms that leverage technological capabilities to transform plaintext messages into ciphertext, effectively encoding the information. Only authorized parties possessing the correct key or secret password can decode the ciphertext back into its original plaintext form [40], [41]. While encryption does not prevent attackers from intercepting the transmitted information, it ensures that the content of the message remains confidential. By making the message unintelligible without the proper key, encryption protects sensitive information from being disclosed, even in the presence of adversarial interception.
2.3.2. Physical layer security
PLS, which is grounded in the principles of information theory, seeks to safeguard wireless communications from eavesdropping by exploiting the intrinsic properties of the physical transmission medium [4]. Unlike conventional encryption methods, PLS operates independently of the computational capabilities of devices. This independence not only ensures robust security but also provides a significant advantage in conserving resources [42], [43]. Furthermore, PLS allows for dynamic adaptation to varying wireless channel conditions by tailoring transmission strategies to the specific characteristics of the physical layer, enhancing its effectiveness in secure communication.
2.3.3. Spread spectrum communication
The spread spectrum technique enhances signal transmission by varying the signal’s frequency through the addition of pseudo-random noise [44], [45]. This process expands the signal’s bandwidth, mitigating the impact of interference, noise, and signal fading. The pseudo-random noise is generated using codes known exclusively to the sender and receiver. At the receiver’s end, these codes are employed to de-spread the signal, allowing the original data to be accurately retrieved. This mechanism not only improves the security of the transmission but also lowers the probability of detection by potential adversaries.
In summary, encryption and PLS primarily focus on protecting the content of the signal, ensuring that unauthorized parties cannot access the transmitted information. On the other hand, both spread spectrum techniques and covert communication aim to obscure the transmission behavior itself. However, while spread spectrum methods enhance transmission security and reduce detectability, they lack a theoretical foundation to ensure covertness.
3. Covert transmission strategies
To effectively avoid detection and interception, various techniques have been cleverly designed to utilize the characteristics of the time, frequency, spatial, power, and modulation domains to achieve covert transmission of information. Fig. 2 categorizes the covert communication research timeline in terms of these five domains, ranging from practical applications of covert communication (e.g., FH and spread spectrum communication techniques) and fundamental information theories (e.g., fast beamforming algorithms and time uncertainty to enhance covertness) to traditional covert communication strategies, which will be specifically discussed in this section, and future intelligent covert communication methods, which will be discussed in detail in Section 4.
3.1. Time domain
Random time-slot selection transmission is the most intuitive type of time-domain covert communication scheme. In fact, time-hopping (TH) techniques were already in use before the square root law was proposed. As shown in Fig. 3, TH techniques not only rely on randomness to counter Willie’s detection but also rely on the complexity of the TH sequence to prevent interception, even if covert communication behavior is detected. Traditional TH communication focuses on enhancing the signal’s anti-jamming and anti-interception capabilities, with less research on the signal concealment performance [7], [8]. Back in 2016, Bash et al. [5] reported the first covert performance analysis of a system in which the communicating party uses a random time-slot selection transmission scheme. Their results showed that the covert rate is exponentially improved compared with the square root law. On this basis, Lu et al. [6] proposed a randomized time-slot selection strategy based on simultaneous consideration of message age and system covertness. Although random time selection has the benefits of both security and covertness, its increased transmission delay requires precise time synchronization.
3.2. Frequency domain
Covert communication techniques in the frequency domain mainly rely on a spread spectrum [11], [12]. It is well known that, when the transmission rate is certain, increasing the signal bandwidth reduces the requirement for the SNR. When the bandwidth is increased to a certain extent, the signal is even allowed to drown under the noise. The spread spectrum technique reduces the signal power density per unit frequency band by spreading the signal energy over a wider frequency band, making it difficult for Willie to recognize it as valid information [9], [10]. Although spread spectrum technology has a strong covert and anti-jamming performance, it is limited by spectrum resources and receiver sensitivity limitations.
As shown in Fig. 3, FH techniques achieve covert communications by rapidly changing the carrier frequency of a signal, making it difficult for a listener to lock onto and track the frequency change of the signal. For example, Torrieri et al. [12] analyzed the performance of FH millimeter-wave uplinks and reported the usefulness of FH for anti-detection, the compensation of frequency-selective fading, and interference reduction. Hui et al. [13] provided a very reliable covert FH transmission scheme, which takes the covertness of the system as the optimization objective and the reliability, average power, and sequence uniformity as the constraints. By jointly optimizing the frequency and power domain coefficients, it can simultaneously achieve the high reliability, anti-interference, and strong covertness of a system.
3.3. Spatial domain
Covert communication technology in the spatial domain mainly realizes covertness by enhancing the isolation of spatial propagation; its technical approaches include beamforming, millimeter wave and terahertz communication, and frequency diverse array (FDA).
3.3.1. Multi-antenna covert communications
Beamforming technology achieves the precise control of the direction of signal transmission by adjusting the phase and amplitude of individual elements of the antenna array to form a beam pointing in a specific direction. This technique allows signals to be transmitted with high intensity in a specific direction while being rapidly attenuated in other directions, greatly reducing signal leakage and the possibility of detection. In Ref. [14], covert communication for both centralized antenna systems (CAS) and distributed antenna systems (DAS) was investigated, with maximal ratio transmitting (MRT) and distributed beamforming (DBF) respectively employed as transmission strategies for CAS and DAS. Furthermore, Wang et al. [15] proposed a robust IRS-assisted beamforming scheme for realizing covert transmission in the space domain. Beamforming can achieve a very high covert rate, but the issue of the efficient alignment of narrow beams remains to be overcome.
3.3.2. Millimeter wave and terahertz covert communications
Combining millimeter wave with beamforming not only improves the rate of covert communications by overcoming severe path loss but also provides covertness in the spatial domain because of its strong directionality [16]. In addition, the higher frequency and short wavelength of millimeter wave make it difficult for signals to penetrate buildings and other objects, which provides an additional level of covertness. However, precisely because of this characteristic, its coverage is severely limited.
The high directivity of the terahertz spectrum also offers advantages for covert communication. Highly directional transmission, which is commonly used in terahertz communications, mitigates the significant path loss caused by severe attenuation from spreading, atmospheric absorption, and scattering effects. This directivity concentrates the transmission power within a narrow beam sector, effectively preventing eavesdroppers outside this sector from successfully detecting the communication. Gao et al. [17] proposed a novel distance-adaptive absorption peak modulation (DA-APM) method for terahertz covert communications. This method enhances covertness by dynamically modulating signals at the molecular absorption peaks of the terahertz spectrum. The results demonstrated that DA-APM can reduce the eavesdropping distance by 60% compared with the random spectrum selection approach, thereby significantly shrinking the insecure region and improving the covertness of terahertz transmissions.
3.3.3. FDA-aided covert communications
As shown in Fig. 4, compared with phased array (PA), in which all array elements operate at the same carrier frequency, FDA technology enables the simultaneous transmission of signals at multiple frequency points by assigning a distinct frequency to each antenna element. This results in a distance- and angle-dependent beam pattern [18]. The ability to shape the beam pattern based on position rather than just direction makes FDA a promising approach for achieving two-dimensional security, and this unique feature has been extensively utilized to enhance covert communications [19]. However, FDA technology is still not mature enough at this time, especially the lack of reliable signal receiver designs.
3.4. Power domain
There are two main methods to achieve covert communication in the power domain: power adaption and AN. For power adaptation schemes, as illustrated in Fig. 5, variations in the transmitted power from the transmitter, jammer, or relay are typically employed [20], [21], [22], [23]. This approach ensures that the received signal power fluctuates randomly, introducing uncertainty for the warden in determining whether the transmitter is actively sending information. However, this method comes with the trade-off of increased complexity in the power amplifier design at the transmitter, jammer, or relay side. For AN schemes, the addition of interference or noise increases the overall uncertainty, encompassing both environmental interference from other users and AN. This allows the system to tolerate stronger transmitted signals [24]. It is worth mentioning that the legitimate receiver will have difficulty detecting the signal, since its received power is also random. Also, by adjusting the power of the transmitted signals, secure transmissions can be concealed within environmental or AN. This noise is typically designed to fluctuate continuously, further reducing the likelihood of detection [25].
3.5. Modulation domain
In the modulation domain, as depicted in Fig. 6, various methods employing random selection schemes, such as digital modulation techniques including multiple-amplitude shift keying (M-ASK), multiple-quadrature amplitude modulation (M-QAM), multiple-phase shift keying (M-PSK), multiple-amplitude and phase shift keying (M-APSK), quadrature phase shift keying (QPSK), and so forth, can be utilized to modulate transmitted signals. These approaches are effective in achieving covert communications [26], [27]. An optimal modulation scheme has been verified to exist for covert communications under specific conditions. On the one hand, low-order modulation outperforms high-order modulation under low SNR conditions because it provides greater reliability, despite its lower information rate. On the other hand, for Willie’s detection, increasing the modulation order results in the received signal statistically following a mixture distribution with more components. According to the central limit theorem, as the number of components increases, this mixture distribution approximates a Gaussian distribution, which typically increases the detection error probability for Willie. Therefore, using higher-order modulation for Alice’s transmission to Bob makes it more challenging for Willie to accurately detect the communication. A covert communication scheme utilizing waveform overlay with weighted fractional Fourier transform (WFRFT) signals has also been proposed [29]. This approach eliminates the need for added noise, improving both the capacity and security of covert communication systems that rely on waveform overlay techniques.
Optimal probabilistic constellation shaping has been explored from a practical perspective as well [28]. Unlike traditional covert communication schemes that use equiprobable constellation modulation, this approach employs non-equiprobable constellations. An approximate gradient descent method was proposed to achieve optimal probabilistic constellation shaping, which further improves the covert rate. The problem of covert rate maximization was addressed by jointly optimizing the constellation distribution and power allocation, enabling more efficient covert communication.
Table 2[5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], [46], [47] summarizes the benefits, limitations, complexities, and application scenarios of typical covert communication techniques in different domains. In addition, some studies focus on multi-domain joint covert communication strategies [46], [47], which are universal and applicable to a wide range of communication scenarios. These strategies have advantages such as strong eavesdropping and interference resistance and high adaptability to complex environments. However, they also present challenges such as complex synchronization designs and difficulties in receiver implementation.
4. Future trends and challenges for intelligent covert communication
Recently, unlike a traditional Willie with a fixed power detection threshold, which may have a higher detection error, the detection accuracy of AI-assisted signal detection techniques has been greatly improved. For example, a novel AI-based integrable detection method demonstrated a significantly improved performance (by 88%–99.9614%) compared with conventional multiple-input and multiple-output (MIMO) detectors in initial simulated MIMO scenarios [48], making it difficult to hide covert communication behaviors. In addition, there is a trend toward intelligent jamming. For example, Amuru et al. [49] used reinforcement learning techniques to enable the jammer to learn the opponent’s communication behavior and adaptively adjust the interference parameters for optimal performance in unknown electromagnetic environments, which will lead to severe limitations in the covert communication rate. Therefore, to satisfy the higher security demands of emerging applications and scenarios, covert communications methodologies require further improvement. In this section, we explore and detail potential future research directions and corresponding challenges.
4.1. Intelligent cooperative covert communication
In current covert communication methods, UAVs with flexible deployment and controlled mobility have been proposed as mobile interference sources to enhance covert communication performance. However, these approaches rely on the mobile interference having prior knowledge of the eavesdropper’s detection threshold, perfect CSI, and noise power, which is often unrealistic. Moreover, eavesdroppers may adopt real-time adjustment strategies, such as dynamically altering their position or detection threshold in response to interference. Existing studies typically assume static eavesdroppers, neglecting these dynamic behaviors. To address the challenges posed by complex and dynamic wireless communication environments, more practical and adaptive ICCC techniques have been developed.
In cooperative covert communication, a generative adversarial network (GAN) approach is utilized to design cooperative jamming strategies that assist unauthorized covert transmissions. For example, the objective functions are configured to maximize both the covert rate and the detection error probability by optimizing the interference power of the jammer. This optimization problem is framed as a dynamic game between the jammer and Alice. To address this game, a GAN-based power optimization algorithm is proposed. The algorithm effectively handles scenarios with limited network environment information and training data, offering a robust solution to achieve a promising trade-off between the covert communication rate and the probability of detection errors. As illustrated in Fig. 7, the generator (a cooperative jammer) takes the environmental state as input and generates interference solution. In turn, the discriminator (Alice) uses the environmental state and the output of G as inputs, and outputs a number D. By comparing D with the predetermined threshold , the GAN judges whether the cooperative-relay is transmitting a covert message. Then, by alternatively training the generator D and discriminator G, the GAN can converge to a Nash equilibrium, achieving an effective trade-off between maximizing the covert communication rate and minimizing the probability of detection errors.
Aside from GANs, ICCCs can also apply the recently popular research trend of deep reinforcement learning (DRL). This dynamic programming approach leverages the ability to learn optimal solutions under changing conditions and adapt to the environment. Existing DRL methods have been applied to various scenarios, such as cyber–physical attack defense and interference management in communication systems [36]. It is anticipated that DRL-based cooperative covert communications will be a significant study area in the future, based on the features of performance optimization in covert communications—that is, the game between legitimate users and eavesdroppers in an imperfect information network.
Nevertheless, for future dynamic wireless channel environments (e.g., different network loads, interference levels, and user mobility), ICCCs need to be robust and adaptive and must dynamically adjust their strategies to achieve optimal covert performance, which requires complex real-time learning algorithms. However, the resulting high computational complexity leads to challenges in the convergence speed and stability of intelligent networks in dynamic environments, so more efficient algorithm design is needed.
4.2. Intelligent parasitic covert communication
Another promising research trend is intelligent parasitic covert communication, which employs spectral superposition to assist the transmitter in achieving covert transmission. More specifically, in a cognitive covert communication network, Alice may use an underlay strategy to transmit her messages. In this scenario, the signal received by the warden is a superimposed signal comprising both legitimate messages and Alice’s unapproved messages. This allows Alice to effectively hide her signals within third-party transmissions. To detect whether Alice is transmitting unapproved signals, the warden performs a binary hypothesis test by comparing the received signal power to a predefined threshold. If the warden fails to make an accurate decision, Alice’s covert transmission is considered successful.
During the process of spectral superposition, inter-system interference arises between the covert and legitimate transmissions. To mitigate this interference and protect legitimate communication, Alice’s transmit power must be constrained. This power limitation, however, reduces the performance of Alice’s covert transmission. To address this challenge, researchers have proposed exploiting improper Gaussian signaling (IGS), which has been shown to improve the degree of freedom and thus outperform conventional proper Gaussian signaling (PGS) schemes in interference-limited networks. The contours of the bivariate Gaussian distribution of the proper Gaussian (PG) random variable (RV) and the improper Gaussian (IG) RV are shown in Fig. 8.
4.3. Multidimensional covert waveform design
Existing covert waveform designs focus on merely one dimension or two dimensions—that is, time, frequency, or both. However, as the electromagnetic environment becomes more complicated and fast changing, it becomes difficult for a communication waveform design based on one or two existing dimensions to fulfill the requirements of strong covertness and high reliability. This motivates the investigation of multidimensional covert waveform design. In particular, the emerging delay-Doppler and affine Fourier domain modulations [31], shown in Fig. 9, can be exploited jointly with the time and frequency domains to facilitate multidimensional covert waveform design. More specifically, the covert channel characteristics can be analyzed to design a joint utilization strategy in the time–frequency, delay-Doppler, and affine Fourier domains. The Fourier transform, Zak transform, and affine Fourier transform are used to generate multi-domain covert waveforms that meet the optimization objective.
It is worth mentioning that designing such a multidimensional covert waveform is not an easy task. Due to the complicated signal processing models, the complexity of the hardware and software design of covert transceivers increases dramatically. In addition, multidimensional covertness constraints often bring about a loss of waveform freedom, which limits the covert rate. In this regard, a communication waveform design based on ML and deep learning models is a promising solution. By continuously inputting environmental information and adjusting the loss function to drive the neural network, the comprehensive communication performance can be optimized and adapted to a dynamic and complex electromagnetic environment for multidimensional covert waveform design. For example, utilizing the characteristics of cooperation between the legitimate transmitter and receiver, as well as the conflict between a malicious warden and the legitimate transceiver, a GAN can be adopted to design a multidimensional covert waveform, where the adversarial neural network is trained to optimize the waveform of the covert signal in terms of the amplitude, phase, and delay-Doppler perspectives. The use of advanced ML and deep learning approaches would require a huge energy consumption due to their iterative process for training and inference. Hence, future works should concentrate on developing energy-efficient AI algorithms with reduced latency and guaranteed covertness.
4.4. Intelligent active detection for covert communication
Unlike typical covert communication scenarios in existing studies, where the eavesdropper is either in a fixed position or performing passive energy detection, the active detection Willie (or active attack Willie) scenario is a typical scenario that must be addressed by future covert communications. A Willie with mobile capability can dynamically change his detection position and power threshold according to the received signal to achieve better detection capability [50]. As an example, a deep-learning-assisted UAV detector can optimize its trajectory, detection threshold, and hovering position in real time with the goal of minimizing the probability of detection error. Moreover, a Willie with attack capability is no longer satisfied with passive detection and can transmit interference signals to actively change the electromagnetic environment when he suspects that Alice is carrying out covert transmission, thereby disturbing and deceiving Alice with sensing capability.
Such intelligent active detectors can compress the viability of covert communications, causing Alice’s judgment of the environmental uncertainty, channel state, and detection ability to be incorrect, and then generate non-credible transmission decision, exposing covert communication behavior. In order to cope with such active detection, the development of Alice’s cognitive capabilities should be considered in the future, including the detection and identification of threats and the credible real-time generation of gaming confrontation strategies.
4.5. ISAC-enhanced intelligent covert communication
With the growing demand for wireless devices, frequency resources are becoming increasingly scarce. In response, ISAC has emerged as a promising 6G technology. ISAC enables wireless systems to simultaneously perform communication and target sensing using a shared waveform, common hardware platform, and unified network infrastructure. The sensing capabilities integrated within ISAC transmission open up novel possibilities for designing secure ISAC-enhanced covert communication. As illustrated in Fig. 10, an Alice with the sensing capability can first sense the presence of a potential Willie and then use the knowledge of the target information to perform active avoidance or introduce more directional interference. More importantly, an integrated sensing and covert communication system can be more effective against an active or mobile Willie, since some levels of the CSI of the covert channel can be probed and exploited with the aid of radar detection.
On the evolutionary path of ISAC-enhanced covert communication design, AI plays a critical role. Mutual interference between communication and sensing waveforms leads to parameter transformations in one system that implicitly reveal information about the other, raising covertness concerns, particularly for military radar. Therefore, it is essential to enable parameter exchange between radar and communication units while preserving the secrecy and covertness of both systems, minimizing mutual interference, and preventing information leakage. Generally, a joint design of sensing and covert communication requires a huge number of parameters that must be optimized and exchanged, such that conventional optimization algorithms are not applicable. To address this challenge, AI-based approaches that leverage the information embedded in the precoder to infer the radar’s location present a promising solution for designing ISAC-enhanced covert communication. This can inspire fruitful research works on deep learning and reinforcement learning techniques to provide efficient and robust solutions. For example, a three-dimensional convolutional neural network (CNN) can be utilized to extract spatial–frequency domain features from the channel frequency response, while a gated recurrent unit is employed to capture the time-domain correlated features of the channel response. A deep reinforcement learning (DRL) framework can also provide an efficient solution of the model parameters. Given the environmental state, the agent selects appropriate actions to maximize the cumulative reward, utilizing the twin delayed deep deterministic policy gradient algorithm to accelerate convergence during training. To enhance covertness, a secure sensing system employs two distinct diffusion models to generate safeguarding signals, which are modulated onto the pilot signals. This masks signal fluctuations caused by user activity, effectively shielding users from unauthorized monitoring.
Nevertheless, from the perspective of information security and covertness, embedding information signaling into the probing waveform for target sensing introduces potential security vulnerabilities. A critical challenge in covert joint sensing communication systems is determining how to design and transmit probing waveforms that embed secure information while avoiding detection by the warden. Moreover, active eavesdroppers with radar sensing capabilities are no less serious a threat.
5. Conclusions
Covert communication is a wireless security technique that provides high-level security by concealing the existence of communication behavior. This work first described the basic concepts of covert communication and compared it with other security techniques such as encryption and PLS. Then, existing research on covert communication was described and compared in detail, using the schemes’ realization domain as a categorization basis. In particular, a clear time line of technology development was shown, and the advantages and disadvantages of covert methods in each domain were summarized and explained. Finally, various applications of intelligent covert communication in future networks and the problems it may present were discussed, and potential research directions were pointed out.
Acknowledgments
This work was supported by the National Natural Science Foundation of China (62425103) and the National Key Research and Development Program of China (2022YFC3301300). The sponsors have no role in the study design, collection, analysis, or interpretation of the data, or in the writing or submission of the manuscript.
Compliance with ethics guidelines
Zan Li, Jia Shi, Jiangbo Si, Lu Lv, Lei Guan, Benjian Hao, Zhuangzhuang Tie, Danyang Wang, Chengwen Xing, and Tony Q.S. Quek declare that they have no conflict of interest or financial conflicts to disclose.
BashBA, GoeckelD, TowsleyD. Limits of reliable communication with low probability of detection on AWGN channels. IEEE J Sel Areas Comm2013; 31(9):1921-1930.
[2]
BashBA, GoeckelD, TowsleyD, GuhaS. Hiding information in noise: fundamental limits of covert wireless communication. IEEE Commun Mag2015; 53(12):26-31.
[3]
HellmanME. An overview of public key cryptography. IEEE Commun Mag2002; 40(5):42-49.
[4]
LiuY, ChenHH, WangL. Physical layer security for next generation wireless networks: theories, technologies, and challenges. IEEE Commun Surv Tutor2017; 19(1):347-376.
[5]
BashB, GoeckelD, TowsleyD. Covert communication gains from adversary’s ignorance of transmission time. IEEE Trans Wirel Commun2016; 15(12):8394-8405.
[6]
LuX, YanS, YangW, LiM, NgDWK. Covert communication with time uncertainty in time-critical wireless networks. IEEE Trans Wirel Commun2023; 22(2):1116-1129.
[7]
DaniV, RamaiyanV, JalihalD. Covert communication over asynchronous channels with timing advantage. In: Proceedings of 2021 IEEE Information Theory Workshop; 2021 Oct 17–21; Kanazawa, Japan; 2021.
[8]
PolydorosA, WeberC. Detection performance considerations for direct-sequence and time-hopping LPI waveforms. IEEE J Sel Areas Commun1985; 3(5):727-744.
[9]
HuY, HanS, ZhaoH, HanY, XuJ, YangG. Extension of the covert performance evaluation system for underwater spread spectrum communication. In: Proceedings of 2023 3rd International Conference on Electronic Information Engineering and Computer (EIECT); 2023 Nov 17–19; Shenzhen, China; 2023.
[10]
DuJ, ZhangY, ChenG. Bionic covert hydroacoustic spread spectrum communication based on improved adaptive filtering algorithm. In: Proceedings of 2024 4th International Conference on Neural Networks, Information and Communication (NNICE); 2024 Jan 19–21; Guangzhou, China. IEEE; 2024. p.1244–9.
[11]
NingB, GuanL, HuangH. A novel frequency-hopping sequence for covert communication. IEEE Access2017; 5:20157-20163.
[12]
TorrieriD, TalaricoS, ValentiMC. Analysis of a frequency-hopping millimeter-wave cellular uplink. IEEE Trans Wirel Commun2016; 15(10):7089-7098.
[13]
HuiP, GuanL, SiJ, LiZ, ZhaoZ, XuR. A power–frequency joint optimization frequency hopping strategy for covert communication. IEEE Commun Lett2023; 27(2):452-456.
[14]
ZhengTX, WangHM, NgDWK, YuanJ. Multi-antenna covert communications in random wireless networks. IEEE Trans Wirel Commun2019; 18(3):1974-1987.
[15]
WangC, LiZ, ShiJ, NgDWK. Intelligent reflecting surface-assisted multi-antenna covert communications: joint active and passive beamforming optimization. IEEE Trans Commun2021; 69(6):3984-4000.
[16]
JamaliMV, MahdavifarH. Covert millimeter-wave communication: design strategies and performance analysis. IEEE Trans Wirel Commun2022; 21(6):3691-3704.
[17]
GaoW, ChenY, HanC, ChenZ. Distance-adaptive absorption peak modulation (DA-APM) for terahertz covert communications. IEEE Trans Wirel Commun2021; 20(3):2064-2077.
[18]
ZhangY, ZhangY, WangJ, XiaoS, TangW. Distance-angle beamforming for covert communications via frequency diverse array: toward two-dimensional covertness. IEEE Trans Wirel Commun2023; 22(12):8559-8574.
[19]
ZhangY, ZhangY, WangJ, XiaoS, NiW, TangW. Robust beamfocusing for FDA-aided near-field covert communications with uncertain location. In: Proceedings of 2023 IEEE International Conference on Communications Workshops (ICC Workshops); 2023 May 28–Jun 1; Rome, Italy. IEEE; 2023. p.1349–54.
[20]
HuJ, YanS, ZhouX, ShuF, LiJ, WangJ. Covert communication achieved by a greedy relay in wireless networks. IEEE Trans Wirel Commun2018; 17(7):4766-4779.
[21]
ShuF, XuT, HuJ, YanS. Delay-constrained covert communications with a full-duplex receiver. IEEE Wirel Commun Lett2019; 8(3):813-816.
[22]
ShahzadK, ZhouX, YanS, HuJ, ShuF, LiJ. Achieving covert wireless communications using a full-duplex receiver. IEEE Trans Wirel Commun2018; 17(12):8517-8530.
[23]
TaoL, YangW, YanS, WuD, GuanX, ChenD. Covert communication in downlink NOMA systems with random transmit power. IEEE Wirel Commun Lett2020; 9(11):2000-2004.
[24]
ShiX, WuD, YueC, WanC, GuanX. Resource allocation for covert communication in D2D content sharing: a matching game approach. IEEE Access2019; 7:72835-72849.
[25]
YangL, YangW, XuS, TangL, HeZ. Achieving covert wireless communications using a full-duplex multi-antenna receiver. In: Proceedings of 2019 IEEE 5th International Conference on Computer and Communications (ICCC); 2019 Dec 6–9; Chengdu, China. IEEE; 2019. p. 912–6.
[26]
YanS, ZhouX, HuJ, HanlySV. Low probability of detection communication: opportunities and challenges. IEEE Wirel Commun2019; 26(5):19-25.
[27]
TopalOA, KurtGK. A countermeasure for traffic analysis attacks: covert communications with digital modulation. IEEE Wirel Commun Lett2021; 10(2):441-445.
[28]
MaS, ZhangY, ShengH, LiH, ShiJ, YangL, et al. Optimal probabilistic constellation shaping for covert communications. IEEE Trans Inf Forensics Security2022; 17:3165-3178.
[29]
MeiL, ShaX, ZhangN. Covert communication based on waveform overlay with weighted fractional Fourier transform signals. In: Proceedings of 2010 IEEE International Conference on Wireless Communications, Networking and Information Security; 2010 Jun 25–27; Beijing, China. IEEE; 2010. p. 472–5.
[30]
ZhouX, YanS, WuQ, ShuF, NgDWK. Intelligent reflecting surface (IRS)-aided covert wireless communications with delay constraint. IEEE Trans Wirel Commun2022; 21(1):532-547.
[31]
HuJ, WuY, ChenR, ShuF, WangJ. Optimal detection of UAV’s transmission with beam sweeping in covert wireless networks. IEEE Trans Vehicular Technol2019; 69(1):1080-1085.
ShiJ, YinZ, TieZ, HuJ, ZhouY, LiuC. Covert performance analysis for OTFS frame structure against sequential change-point detection. In: Proceedings of 2023 International Conference on Ubiquitous Communication; 2023 Jul 7–9; Xi'an, China. IEEE; 2023. p. 209–14.
[36]
LuongNC, HoangDT, GongS, NiyatoD, WangP, LiangYC, et al. Applications of deep reinforcement learning in communications and networking: a survey. IEEE Commun Surv Tutor2019; 21(4):3133-3174.
[37]
LiZ, ShiJ, WangC, WangD, LiX, LiaoX. Intelligent covert communication design for cooperative cognitive radio network. China Commun2023; 20(7):122-136.
[38]
HuJ, LinQ, YanS, ZhouX, ChenY, ShuF. Covert transmission via integrated sensing and communication systems. IEEE Trans Vehicular Technol2024; 73(3):4441-4446.
[39]
HuangKW, WangHM, TowsleyD, PoorHV. LPD communication: a sequential change-point detection perspective. IEEE Trans Commun2020; 68(4):2474-2490.
[40]
BhanotR, HansR. A review and comparative analysis of various encryption algorithms. Int J Secur Its Appl2015; 9(4):289-306.
[41]
AleneziMN, AlabdulrazzaqH, MohammadNQ. Symmetric encryption algorithms: review and evaluation study. Int J Commun Netw Inf Secur2020; 12(2):256-272.
[42]
WangD, BaiB, ZhaoW, HanZ. A survey of optimization approaches for wireless physical layer security. IEEE Commun Surv Tutor2019; 21(2):1878-1911.
[43]
SunL, DuQ. Physical layer security with its applications in 5G networks: a review. China Commun2017; 14(12):1-14.
[44]
PickholtzRL, MilsteinLB, SchillingDL. Spread spectrum for mobile communications. IEEE Trans Vehicular Technol1991; 40(2):313-322.
[45]
RamaniV, SharmaSK. Cognitive radios: a survey on spectrum sensing, security and spectrum handoff. China Commun2017; 14(11):185-208.
[46]
LiK, SobersTV, TowsleyD, GoeckelD. Covert communication in continuous-time systems in the presence of a jammer. IEEE Trans Wirel Commun2022; 21(7):4883-4897.
[47]
BaiL, XuJ, ZhouL. Covert communication for spatially sparse mmWave massive MIMO channels. IEEE Trans Commun2023; 71(3):1615-1630.
[48]
DahaMY, RaffertyJ, AshrafMI, HadiMU. Artificial intelligence-enhanced signal detection technique for beyond fifth generation networks. Comput Electr Eng2024; 119:109608.
[49]
AmuruSD, TekinC, vanM der Schaar, BuehrerRM. Jamming bandits—a novel learning method for optimal jamming. IEEE Trans Wirel Commun2016; 15(4):2792-2808.
[50]
ZhaoY, LiZ, ChengN, WangW, LiC, ShenX. Covert localization in wireless networks: feasibility and performance analysis. IEEE Trans Wirel Commun2020; 19(10):6549-6563.
PDF
(2155KB)
