Wireless physical layer information hiding method and system based on neural network

Abstract

The invention discloses a wireless physical layer information hiding method and system based on a neural network, and belongs to the technical field of wireless communication security. The method realizes information hiding by modifying a physical layer waveform of a wireless signal, and specifically comprises the following steps: (1) transmitter secret information embedding: a transmitting end introduces imperceptible tiny disturbance into the physical layer waveform by using a pre-trained neural network encoder, and the secret information is embedded into the physical layer waveform; and (2) transmitter secret information extraction: a receiving end extracts secret information from the received physical layer signals through the paired neural network decoders. The method is a general physical layer information hiding technology, is suitable for various wireless protocols and signal modulation modes such as LoRa, Bluetooth Low Entry (BLE), Wi-Fi, NB-IoT and the like, and can be widely applied to transmitter fingerprint identification, equipment authentication, intellectual property protection (watermarking), covert communication and the like. The present specification will provide two embodiments of BLE and LoRa.

Description

Technical Field

[0001] The present invention relates to a wireless physical layer information hiding method and system based on neural network, belonging to the field of wireless physical layer security. Background Art

[0002] Physical layer (PHY) information hiding is a novel approach to secure communications by directly embedding secret information within the physical properties of wireless signals. This technology has important applications in transmitter fingerprinting, device authentication, intellectual property protection, and the construction of covert communication channels. Currently, research on PHY information hiding falls into two main categories. The first focuses on theoretical analysis and the construction of covert transmission schemes based on idealized models. For example, some studies utilize multi-antenna technology, reconfigurable smart surfaces (RIS), or millimeter-wave communication technologies to achieve directional transmission. Through precise beamforming, signal energy is focused in the direction of the legitimate receiver, minimizing signal leakage to potential eavesdroppers in other directions. Other studies explore superimposing carefully designed artificial noise within legitimate communication signals, enabling legitimate receivers to decode the signals while making it difficult for eavesdroppers to detect the presence of the communication due to insufficient signal-to-noise ratio. However, these theoretical approaches often rely on complex algorithmic design, accurate knowledge of channel state information (especially eavesdropper channel information), or high hardware platform requirements (such as large-scale antenna arrays and high-performance RIS units). These idealized assumptions make them difficult to apply in practical deployments. The second category of research focuses more on system-level implementation and hardware prototyping, often embedding secret information by customizing the modulation schemes or signal parameters of existing wireless communication protocols. For example, one study, based on the LoRa protocol, hides additional data by fine-tuning the amplitude or phase of its linear frequency modulation (chirp) signal. Another example is the use of slightly shifting the positions of constellation points in the I / Q constellation diagram to embed radio frequency watermarks in narrowband Internet of Things (NB-IoT) systems. While these methods have been experimentally validated for their feasibility and effectiveness under specific wireless protocols, they generally suffer from a high degree of protocol dependency. Covert communication schemes designed for a specific protocol (such as LoRa) are often difficult to directly migrate to other protocols that utilize different physical layer mechanisms (such as Bluetooth Low Energy, Wi-Fi, or 5G New Radio), significantly limiting their versatility and applicability. Therefore, developing a versatile physical layer information hiding framework that can span multiple wireless protocols remains a key research direction in this field.

[0003] Meanwhile, in the field of digital media information hiding (including steganography and digital watermarking in images, video, text, and audio), the introduction of deep learning, particularly neural network-based encoder-decoder architectures, has fundamentally transformed the traditional technical paradigm, which relies on artificial feature design or specialized transform domain modifications. Through end-to-end training, these architectures enable neural networks to automatically learn optimal secret information embedding strategies and robust extraction methods. This balance between hiding capacity, imperceptibility, and robustness against various attacks (such as compression and noise) is achieved, significantly improving overall performance and becoming a mainstream approach in this field. However, despite the significant achievements of deep learning in these digital media information hiding areas, its application to physical layer information hiding in wireless communications, particularly encoder-decoder architectures that embed and extract secret information directly based on RF waveforms, remains a research gap that has not been fully explored. Research in this area is expected to overcome the limitations of existing technologies in terms of protocol dependence, versatility, and practicality, becoming a key breakthrough in promoting the development of physical layer information hiding technology.

[0004] Directly migrating well-established deep learning information hiding methods from digital media to the physical layer of wireless communications presents a series of unique and critical technical challenges. First, wireless signals are fundamentally different from precisely controllable digital carriers like images or text. In complex electromagnetic propagation environments, signals transmitted through air channels are inevitably subject to random variations in signal amplitude and phase (i.e., fading) caused by multipath effects and shadow fading, as well as interference from various environmental noise sources. Furthermore, transceiver hardware itself exhibits numerous non-idealities, such as nonlinear distortion in power amplifiers, phase noise in oscillators, and imbalance in I / Q channels, all of which can cause significant distortion to the RF waveform. Ensuring that embedded, weak secret information can survive and be accurately extracted under such complex and uncontrollable time-varying channel conditions and hardware impairments is a significant test of the robustness of neural network models. Second, and more critically, any physical-layer information hiding technique must operate without interfering with, or only negligibly impacting, the normal functioning of the primary communication link. This means that the signal perturbations introduced by the secret information embedding process must be sufficiently subtle so as not to significantly degrade the demodulation performance of the primary communication data portion or the overall communication quality. Therefore, how to embed the covert information while ensuring the reliability, transmission efficiency, and protocol compatibility of the primary communication constitutes a core contradiction and technical bottleneck that must be prioritized in designing a practical physical layer information hiding system. Summary of the Invention

[0005] In order to solve the above problems, the present invention proposes a wireless physical layer information hiding method and system based on neural network, which directly operates the physical layer waveform through end-to-end training of neural network, breaking through the limitations of existing technologies.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A wireless physical layer information hiding method based on a neural network comprises the following steps:

[0008] (1) Build and train a neural network-based secret information embedder and extractor in a controlled environment;

[0009] (2) deploying the secret information embedder and extractor at the transmitter and receiver respectively;

[0010] (3) The transmitter uses the secret information embedder to introduce waveform-level weak distortion into the wireless physical layer waveform to embed the secret information, and transmits the distorted signal into the air;

[0011] (4) After the receiver captures the wireless physical layer signal, the extractor is used to extract the secret information from the waveform;

[0012] The distortion introduced by the secret information embedder is limited to have no significant impact on the demodulation performance of the main communication link.

[0013] Preferably, the secret information embedder and extractor in step (1) are trained using an end-to-end training strategy. During the training process, additive white noise, multipath fading, and frequency offset channel conditions are simulated using a wireless channel simulator, and the weights of the secret information embedder and extractor are synchronously updated using a backpropagation algorithm.

[0014] Preferably, the secret information embedder in step (3) is implemented using a deep neural network, with the input being the secret information and the standard physical layer waveform, and the output being a distorted waveform containing the secret information.

[0015] Preferably, the method does not rely on a specific communication protocol and is applicable to the physical layer waveforms of LoRa, Bluetooth Low Energy, Wi-Fi and NB-IoT.

[0016] Preferably, the method further comprises calibrating the secret information embedder and extractor using measured physical layer waveform data in a real hardware environment to adapt to actual channel fading and hardware nonlinear distortion.

[0017] Preferably, after extracting the secret information in step (4), forward error correction decoding and deinterleaving operations are performed to recover the original secret information, and the forward error correction code includes BCH or LDPC coding technology.

[0018] Preferably, the implementation platform of the transmitter and receiver is a software-defined radio or a radio frequency integrated circuit, and the software-defined radio supports real-time operation of physical layer waveforms and neural network deployment.

[0019] Preferably, a recovery loss and a concealment loss are defined during the training process, wherein the recovery loss constrains the extraction accuracy of the secret information through a mean square error function, and the concealment loss limits the degree of waveform distortion through a mean square error function.

[0020] Preferably, the forward error correction decoding and deinterleaving operations are dynamically enabled according to channel conditions.

[0021] The present invention also provides a wireless physical layer information hiding system based on the method, comprising:

[0022] Transmitter module: used to generate a physical layer waveform containing concealed distortion through a secret information embedder;

[0023] Receiver module: used to recover hidden information through secret information extractor;

[0024] Channel simulator module: integrated into the training phase to simulate additive white noise, multipath fading, and frequency offset channel conditions;

[0025] Forward Error Correction (FEC) module: deployed in the transmitter and receiver, supporting BCH or LDPC encoding and decoding.

[0026] The innovative features of the present invention include:

[0027] (1) Secret information embedding method: At the transmitter, the secret information is mapped into scrambled

[0028] This disturbance affects the normal synchronization and demodulation process at the receiving end.

[0029] There is no obvious impact, thus achieving covert secret data embedding.

[0030] (2) Secret information extraction method: At the receiving end, a paired neural network decoder is used to decode the received packet.

[0031] The wireless physical layer waveform containing disturbance is analyzed to recover the embedded secret information.

[0032] (3) Neural network training and fine-tuning in a controlled environment: An end-to-end strategy is used to train the secret embedding network and the secret extraction network, and the model is optimized under channel simulation conditions such as additive white noise, multipath fading, and frequency offset.

[0033] In view of the differences in actual hardware and channels, the present invention also introduces a fine-tuning mechanism to further tune the model parameters in the real channel and hardware environment, significantly improving the performance and robustness of the system in the real environment.

[0034] In summary, the physical layer information hiding method proposed in this invention uses a deep neural network to covertly embed secret information into the wireless physical layer waveform, which can be widely used in applications such as transmitter fingerprint recognition, device authentication, intellectual property protection (watermarking), and covert communication.

[0035] Beneficial Effects: The physical layer information hiding method provided by the present invention uses a deep neural network to embed weak distortion into the physical layer waveform of a wireless signal to hide secret information. The present invention innovatively uses a deep neural network for physical layer information hiding tasks, and its advantages over existing technologies are as follows:

[0036] (1) Deep neural networks are used to embed distortion into wireless physical layer signals, effectively increasing the embedding capacity and concealment of secret information.

[0037] (2) The invention directly embeds distortion into the physical layer waveform (I / Q samples), is universal, and can be applied to any wireless protocol and signal format. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of wireless physical layer information hiding system.

[0039] Figure 2 Flowchart of the wireless physical layer information hiding system based on deep learning.

[0040] Figure 3 Training process for secret information embedder and secret information extractor.

[0041] Figure 4 An example of BLE physical layer waveform embedding, where (a) is the original standard physical layer waveform, and (b) and (c) are the physical layer waveforms after embedding hidden information of different lengths.

[0042] Figure 5 An example of BLE physical layer waveform embedding, where (a) is the original standard physical layer waveform, and (b) and (c) are the physical layer waveforms after embedding hidden information of different lengths.

[0043] Figure 6 On-site photos of the wireless physical layer information hiding method based on deep learning, where (a) is an indoor environment and (b) is an outdoor environment. DETAILED DESCRIPTION

[0044] 1. System Overview

[0045] like Figure 2 As shown in the figure, the wireless physical layer information hiding method based on deep learning proposed in the present invention includes two communicating parties, namely transmitter Alice and receiver Bob. They have established a primary communication link, such as Wi-Fi, LoRa, LTE, etc. The purpose of physical layer information hiding is to embed additional hidden information in the transmitted wireless data packets. In more detail, the sender Alice introduces slight distortion by using a pre-trained neural network (called a secret information embedder) to hide the secret information in the physical layer waveform. The distortion introduced by this information embedding process is very weak, so that the main communication channel is not affected. After receiving the signal, Bob uses another neural network called a "secret information extractor" to extract the secret information. In addition, depending on the channel conditions, the system can also use forward error correction code technology to correct transmission errors and improve the reliability of hidden information transmission. The specific operations of transmitter Alice and receiver Bob are described below.

[0046] (1) Transmitter operation

[0047] Alice first performs forward error correction encoding and interleaving to convert the secret information into a codeword. A pre-trained neural network secret information embedder then introduces a subtle and covert waveform-level distortion to embed the codeword into the standard waveform. Finally, this distorted signal is upconverted from baseband to the RF band and transmitted into the air via an antenna.

[0048] (2) Receiver operation

[0049] Bob captures the RF signal and downconverts it to baseband. He then uses a pre-trained neural network secret information extractor to extract the codeword hidden in the physical layer waveform. The extracted codeword is then fed into a deinterleaver and forward error correction decoder for error correction, thereby recovering the secret information sent by Alice.

[0050] It is worth noting that the two neural networks, the secret information embedder and extractor, are pre-trained in a controlled environment and then secretly shared by Alice and Bob. This physical layer information hiding method can be implemented using software-defined radios or customized radio frequency integrated circuits.

[0051] 2. Transmitter secret information embedding process

[0052] The goal of transmitter Alice is to embed secret information into wireless data packets by introducing waveform-level distortion without affecting the quality of the primary communication link. This can be achieved through the following three modules.

[0053] (1) Forward error correction code encoding and interleaving

[0054] When the channel transmission quality is unstable, Alice first performs forward error correction coding to introduce redundancy, and then uses a random interleaver to rearrange the elements in order to mitigate the impact of burst errors that may occur during transmission. Coding technologies such as BCH and LDPC are preferred.

[0055] (2) Secret Information Embedder

[0056] After converting the secret information into codewords, the transmitter uses a pretrained neural network secret information embedder to hide it within the physical layer waveform. This neural network takes as input the codeword to be embedded and a standard physical layer waveform, and outputs a distorted waveform containing the secret information. In a preferred embodiment, the secret information embedder can be constructed by stacking a large number of convolutional layers, pooling layers, and linear layers; other neural network structures are alternatives.

[0057] 3. Receiver secret information extraction process

[0058] The radio signal transmitted by Alice propagates through the wireless channel and is then captured by Bob. It is down-converted to baseband and then passes through the following modules for secret information extraction.

[0059] (1) Secret Information Extractor

[0060] Bob first uses a pre-trained secret information extractor to recover the information hidden in the physical layer waveform. Its input is the received physical layer waveform, and its output is the extracted codeword. In the preferred embodiment, the bit extractor structure can be composed of a large number of stacked convolutional layers, pooling layers, and linear layers. Other neural network structures are possible alternatives.

[0061] (2) Deinterleaving and forward error correction code decoding

[0062] After extracting the codewords from the physical layer waveform, Bob uses a random deinterleaver to reconstruct their original order and then uses a forward error correction code decoder to recover the secret message that Alice intended to send.

[0063] 4. Secret Information Embedder and Extractor Training Process

[0064] Neural network-based secret information embedder and extractor are the basic components of the proposed physical-layer information hiding system, which are pre-trained in a controlled simulation environment and then shared by Alice and Bob.

[0065] The joint training process of the secret information embedder and extractor proposed in this method is as follows Figure 3As shown in the figure. As can be seen, this method simulates the communication process between Alice and Bob through a secret information embedder, an extractor, and a wireless channel simulator. At the beginning of each training step, a set of binary vectors are randomly generated. More specifically, each element in the binary vector is randomly assigned to 0 and 1, that is, uniformly distributed. Then, the secret information is hidden in a standard waveform using a bit embedder to generate a distorted signal. Subsequently, the distorted signal is input into the wireless channel simulator. In particular, it can simulate channel fading and add noise. Finally, the output of the wireless channel simulator is fed into the secret information extractor to recover the embedded ciphertext.

[0066] To constrain the neural network training process, the present invention defines two loss functions: recovery loss and concealment loss. The purpose of the recovery loss is to ensure that secret information can be successfully embedded and extracted. The concealment loss limits the degree of waveform distortion, making the covert link undetectable to eavesdroppers. In the preferred embodiment of the present invention, the recovery loss and concealment loss are implemented using the mean square error function. Other functions are alternatives. The root mean square function is expressed as follows:

[0067]

[0068] Among them, L rec To recover the loss, c and are the secret information embedded by the transmitter and recovered by the receiver, and N is the length of the secret information. ste is the hidden loss, w and are the standard signal waveform and the wireless signal waveform after embedding distortion, and M is the length of the discrete waveform.

[0069] After training in a simulation environment, the system can further calibrate the neural network weights using measured physical layer signals collected from an actual hardware platform. This process, known as fine-tuning, optimizes model performance, adapting it to fading and hardware impairments in real-world wireless channels, and bridging the gap between simulation and reality.

[0070] 5. Implementation and Prototype Platform

[0071] This method embeds and extracts information directly from the physical layer waveform, demonstrating its versatility and applicability across a wide range of communication protocols and signal formats. To demonstrate the applicability and feasibility of this method, this section examines Bluetooth Low Energy (BLE) and LoRa as two representative implementations and describes their corresponding prototype platform implementations.

[0072] (1) Example 1: Physical layer information hiding system for Bluetooth Low Energy (BLE) protocol

[0073] The BLE protocol uses Gaussian frequency shift keying (GFSK) modulation at the physical layer. The transmitter adds redundancy to the secret information using BCH forward error correction (FEC) code and uses a neural network embedder to hide it within the standard BLE waveform, creating a distorted BLE signal with the secret information and transmitting it. After the receiver captures the physical layer waveform, it feeds the secret information extractor to recover the codeword, which is then fed into the FEC decoder and deinterleaver to recover the secret information intended by the transmitter.

[0074] Figure 4 The standard BLE waveform and the waveform after introducing distortion are shown. Note that by imposing constraints on the neural network training process, the introduced distortion is very weak and does not affect the normal demodulation function of the BLE data packet itself, and the payload part can still transmit data normally.

[0075] (2) Example 2: Physical layer information hiding system for LoRa protocol

[0076] The LoRa protocol uses chirped spread spectrum (CSS) modulation at its physical layer. The physical layer waveform typically consists of a series of identical rising chirps. The transmitter uses BCH forward error correction (FEC) code to add redundancy to the secret information and uses a neural network embedder to hide it within the standard LoRa waveform, creating a distorted LoRa signal with the secret information and transmitting it. After the receiver captures the physical layer waveform, it feeds the secret information extractor to recover the codeword, and then feeds the FEC code decoder and deinterleaver to recover the secret information intended by the transmitter.

[0077] Figure 5 The standard LoRa waveform and the waveform after introducing distortion are shown. Note that by imposing constraints on the neural network training process, the introduced distortion is very weak and does not affect the normal demodulation function of the LoRa data packet itself, and the payload part can still transmit data normally.

[0078] Prototype platform

[0079] In order to successfully implement the wireless physical layer information hiding technology proposed in the present invention, the transmitting and receiving devices must have the ability to run neural networks and allow access to and operation of physical layer waveforms. At the same time, these devices must also have the function of transmitting customized RF waveforms. Software-defined radio technology, by decoupling software from hardware, provides extremely high flexibility in digital signal processing, which fully meets the above deployment requirements. Software-defined radio technology has been widely used in many fields such as satellite communications, national defense, drone systems, and 5G cellular networks, which greatly reduces the implementation threshold of deep learning-based RF steganography technology. It is worth mentioning that although software-defined radio is the preferred implementation platform of the present invention, any other platform that can realize neural network signal processing at the physical layer, such as the joint / integrated development of artificial intelligence chips and RF chips, is also included in the protection scope of the present invention.

[0080] like Figure 6 As shown in Figure 1, the prototype platform uses two software-defined radio platforms as wireless transceivers, which are connected to two running laptops to perform wireless physical layer information hiding tasks.

Claims

1. A wireless physical layer information hiding method based on neural network, characterized in that: The steps include: (1) Build and train a neural network-based secret information embedder and extractor in a controlled environment; (2) deploying the secret information embedder and extractor at the transmitter and receiver respectively; (3) The transmitter uses the secret information embedder to introduce waveform-level weak distortion into the wireless physical layer waveform to embed the secret information, and transmits the distorted signal into the air; (4) After the receiver captures the wireless physical layer signal, the extractor is used to extract the secret information from the waveform; The distortion introduced by the secret information embedder is limited to have no significant impact on the demodulation performance of the main communication link.

2. The method according to claim 1, characterized in that The secret information embedder and extractor described in step (1) are trained through an end-to-end training strategy. During the training process, the wireless channel simulator is used to simulate additive white noise, multipath fading and frequency offset channel conditions, and the weights of the secret information embedder and extractor are synchronously updated through the back propagation algorithm.

3. The method according to claim 1, characterized in that The secret information embedder in step (3) is implemented using a deep neural network, with the input being the secret information and the standard physical layer waveform, and the output being a distorted waveform containing the secret information.

4. The method according to claim 1, wherein The method is independent of specific communication protocols and is applicable to the physical layer waveforms of LoRa, Bluetooth Low Energy, Wi-Fi, and NB-IoT.

5. The method according to claim 1, wherein The method further includes calibrating the secret information embedder and extractor using measured physical layer waveform data in a real hardware environment to adapt to actual channel fading and hardware nonlinear distortion.

6. The method according to claim 1, characterized in that After extracting the secret information in step (4), forward error correction decoding and deinterleaving operations are performed to recover the original secret information, wherein the forward error correction code includes BCH or LDPC coding technology.

7. The method according to claim 1, characterized in that The implementation platform of the transmitter and receiver is software-defined radio or radio frequency integrated circuit, and the software-defined radio supports real-time operation of physical layer waveforms and neural network deployment.

8. The method according to claim 2, characterized in that During the training process, a recovery loss and a concealment loss are defined. The recovery loss constrains the extraction accuracy of secret information through a mean square error function, and the concealment loss limits the degree of waveform distortion through a mean square error function.

9. The method according to claim 6, characterized in that The forward error correction decoding and deinterleaving operations are dynamically enabled based on channel conditions.

10. A wireless physical layer information hiding system based on the method according to any one of claims 1 to 9, characterized in that: include: Transmitter module: used to generate a physical layer waveform containing concealed distortion through a secret information embedder; Receiver module: used to recover hidden information through secret information extractor; Channel simulator module: integrated into the training phase to simulate additive white noise, multipath fading, and frequency offset channel conditions; Forward Error Correction (FEC) module: deployed in the transmitter and receiver, supporting BCH or LDPC encoding and decoding.

Images