Vortex optical communication method and system based on deep neural network

Abstract

The invention discloses a vortex optical communication method and system based on a deep neural network. The method comprises the following steps: constructing a deep neural network initial model and training to obtain a prediction model; the transmitting end encrypts information through a code table to generate a vortex light beam, and the receiving end receives an intensity distribution picture of the vortex light beam and decrypts the intensity distribution picture to obtain information; generating a ciphertext string based on the current Unix timestamp and the private key, acquiring the same byte from the ciphertext string by the transmitting end and the receiving end, generating a vortex light beam carrying the topological charge corresponding to the byte by the transmitting end, receiving and storing an intensity distribution picture corresponding to the vortex light beam by the receiving end, and marking the acquired byte as a picture label; and predicting bytes of the intensity distribution picture by utilizing a prediction model, judging whether the mark is correct or not according to a prediction result, further training the prediction model by utilizing a new training set, obtaining a trained model to replace the initial model, and repeating the process to realize safe information transmission. According to the method, the neural network, encryption and vortex optical communication technologies are combined, the optimization model is continuously trained in actual communication, and the prediction accuracy and the information transmission safety are improved.

Description

Technical Field

[0001] The present invention relates to the field of optical communication technology, and in particular to a vortex optical communication method and system based on deep neural network. Background Art

[0002] Vortex beam (VB) is a type of beam with unique properties, in which each photon carries The orbital angular momentum (OAM) of the vortex beam is spiral, and its phase wavefront function can be used Expressed as, where l is the topological charge, is the chirality coordinate. Theoretically, the topological charge of a vortex beam can take on any integer value, and in this case, different vortex beams are orthogonal to each other. Based on this property, vortex beams carrying different integer topological charges can be simultaneously transmitted coaxially, achieving orbital angular momentum multiplexing (OAM-DM), which significantly increases channel capacity without introducing crosstalk during transmission.

[0003] However, when vortex beams propagate in an atmospheric environment, the orthogonality between different vortex beams is destroyed due to the interference of atmospheric turbulence, which in turn causes crosstalk. In addition, vortex beams propagating in free space are easily split and copied by others, putting the information carried by the beams at risk of being cracked. In view of this, in-depth research on how to build a vortex optical communication system with self-learning and adaptive capabilities and resistance to cracking is of vital importance for the subsequent application and expansion of vortex beam research.

[0004] Current vortex beam communication methods are primarily based on grating diffraction separation and deep convolutional neural network interpretation. The grating diffraction separation method separates the different modes of light in a vortex beam into different spatial coordinates, forming a pattern intensity map. This pattern intensity map is captured by a camera and fed into a trained deep convolutional neural network for interpretation, thereby extracting the information carried by the vortex beam and enabling vortex beam communication. However, this method requires the active collection and labeling of large amounts of data to train the deep convolutional neural network. Furthermore, the accuracy of recognition can be severely affected by skewed or off-axis positioning of the receiving end, camera defects, and contaminated mirrors. Furthermore, the transmitted vortex beam is easily split and replicated, and the original data is demodulated through semantic analysis.

[0005] For example, the document "Turbo-coded 16-ary OAM shift keying FSO communication system combining the CNN-based adaptive demodulator" (Optics Express 26, 27849-27864 (2018)) discloses a vortex optical communication method based on a deep convolutional neural network. This method first turbo encodes the binary information sequence, and then maps the encoded sequence to different OAM mode values through the SLM. After free space transmission, the receiver captures the light intensity image and sends it to the trained CNN network to realize OAM beam demodulation. Finally, turbo decoding and sequence demapping are performed to restore the transmitted information. Although this method can combat the impact of turbulence on the light beam to a certain extent, it requires manual labeling of a large amount of data for deep convolutional neural network learning, and this method does not consider the impact of the receiver's position deviation, off-axis, camera bad pixels, and mirror group dirt on recognition accuracy. In addition, the transmitted vortex beam can be easily split and copied, and the original data is demodulated through semantic analysis.

[0006] The patent, entitled "A multi-mode vortex beam demultiplexing method based on neural network", with patent number CN202110028012.5, constructs a simulation data set, designs a deep convolutional neural network with adjustable parameters, modifies the adjustable parameters of the model according to the training effect, optimizes the model performance, and completes the vortex beam mode demodulation. This demultiplexing method can perform better in more complex multi-mode coupling situations. It can parse the amplitude of each mode based on the input image and improve the effect of optical communication. However, the process of modifying the adjustable parameters of the model in this method is complicated and labor-intensive, and the method does not consider the situation where the transmitted vortex beam may be copied and cracked. Summary of the Invention

[0007] Purpose of the invention: The purpose of the present invention is to provide a vortex light communication method and system based on deep neural network, which can accurately identify the information carried by vortex light and has anti-cracking capabilities.

[0008] Technical solution: To achieve the above objectives, the present invention provides a vortex optical communication method based on a deep neural network, wherein the transmitting and receiving ends share the same private key and the same time system, including the following steps:

[0009] S1. Build an initial deep neural network model, train the initial model using the training set, and obtain an initial prediction model;

[0010] S2: The sending end encrypts the information T through the code table A and generates a vortex beam to send. The receiving end receives the intensity distribution image P corresponding to the vortex beam. t , and use the initial prediction model to predict the intensity distribution image P t Make an interpretation and obtain information T;

[0011] S3. After the information T is sent, the sender and receiver generate a ciphertext string based on the current Unix timestamp and private key, and obtain a byte B from the ciphertext string; the sender encrypts byte B using code table B and generates a vortex beam to send, and the receiver obtains the intensity distribution image P corresponding to the vortex beam. b , and use the obtained byte B as the intensity distribution picture P b byte labels, the intensity distribution image P b and labels are stored in a new training set;

[0012] S4. Using the initial prediction model to analyze the intensity distribution image P b The corresponding byte is predicted. If the predicted byte is the same as the byte B obtained by the receiving end, the prediction is judged to be correct. If they are different, the prediction is judged to be wrong. Based on the judgment result, the initial prediction model is further trained with the new training set to obtain the trained prediction model, and the initial prediction model in S1 is replaced;

[0013] S5. Repeat S2-S4 to achieve secure information transmission between the sending and receiving ends.

[0014] The method for obtaining the initial prediction model described in S1 is:

[0015] S101. Both the sending and receiving ends obtain the current Unix timestamp, encrypt the private key and the Unix timestamp to generate a l-bit ciphertext string, take the low n bits or high n bits of the ciphertext string as an offset, and obtain an m-bit byte A starting from the offset. <l;

[0016] S102: The transmitting end encrypts the acquired byte A by querying the code table B and generates a vortex beam to be sent; the receiving end obtains the intensity distribution image P corresponding to the vortex beam after passing through the Dammann grating and focusing lens. a The receiving end uses the obtained byte A as the intensity distribution picture P a Byte tag of

[0017] S103. Repeat steps S101-S102 until a certain number of intensity distribution images with byte labels are obtained to construct an initial training set, train the deep neural network model based on the initial training set, learn the relationship between the labels and the intensity distribution images, and obtain an initial prediction model.

[0018] The code table A includes a mapping relationship between information and bytes, and the code table B includes a mapping relationship between bytes and topological charges.

[0019] Among them, the transmitting end generates a vortex beam carrying a topological charge based on the topological charge, and the receiving end obtains the intensity distribution picture corresponding to the vortex beam by generating a hologram about a specific topological charge and loading it onto a digital micromirror array. After parallel light illuminates the digital micromirror array, a vortex beam carrying the topological charge is generated. The vortex beam passes through a Dammann grating and a lens and is focused onto a camera, thereby obtaining a picture of the intensity distribution.

[0020] Among them, the method for the sender to encrypt and send the information T in S2 and the receiver to decrypt and obtain the information T after receiving it is as follows: the sender converts the actual information T to be sent into bytes in ciphertext form by querying the code table A, and the sender converts the bytes in ciphertext form into the corresponding topological charge number, generates the corresponding vortex beam based on the topological charge number, and obtains the intensity distribution image P corresponding to the vortex beam at the receiver. t , using the initial prediction model to analyze the intensity distribution image P t Make an interpretation and output the intensity distribution picture P t The corresponding bytes are the information that the sender actually wants to send.

[0021] Among them, in S4, the initial prediction model is further trained using the new training set based on the judgment result to obtain the trained prediction model as follows:

[0022] S401, after a certain number of determinations, calculate the accuracy of the initial prediction model; if the accuracy is less than the set threshold, retrain the initial model using a new training set, and stop when the loss function converges;

[0023] S402: If the accuracy is greater than the set accuracy and the number of intensity distribution images stored in the new training set is greater than the set number, the initial prediction model is retrained using the new training set, and the trained model is used to replace the initial prediction model in S1.

[0024] Among them, the information transmission between the sending end and the receiving end is based on the data packet as the basic unit, including the packet header bit, data bit, identification bit, training data bit, and packet tail bit. If the receiving end does not receive the packet header during the process of receiving the data packet, it will continue to wait for the packet header. If the packet header is received, it will start to receive the data bits. When at least four data bits are received, it will start to wait for the identification bit. If the identification is not received, it will continue to receive data. If the identification is received, the data reception is completed, and it will start to receive training data and wait for the packet tail bit. If the packet tail is not received, it will continue to receive data. If the packet tail is received, the entire data packet is received.

[0025] Among them, the data bits are the ciphertext of the information to be transmitted, and the actual information can be directly output after being interpreted by the neural network at the receiving end.

[0026] The training data bits following the identification bits are used to collect training data.

[0027] The present invention provides a vortex optical communication system based on a deep neural network, wherein the transmitting and receiving ends share the same private key and the same time system, and includes the following modules:

[0028] Initial prediction model generation module: builds the initial model of the deep neural network, trains the initial model using the training set, and obtains the initial prediction model;

[0029] Information transmission module: The sending end encrypts the information T through the code table A and generates a vortex beam to send, and the receiving end receives the intensity distribution image P corresponding to the vortex beam t , and use the initial prediction model to predict the intensity distribution image P t Make an interpretation and obtain information T;

[0030] Training set generation module: After the information T is sent, the sender and receiver generate a ciphertext string based on the current Unix timestamp and private key, and obtain a byte B from the ciphertext string; the sender encrypts byte B through code table B and generates a vortex beam to send, and the receiver obtains the intensity distribution image P corresponding to the vortex beam b , and use the obtained byte B as the intensity distribution picture P b byte labels, the intensity distribution image P b and labels are stored in a new training set;

[0031] Initial prediction model training module: Use the initial prediction model to train the intensity distribution image P b The corresponding byte is predicted. If the predicted byte is the same as the byte B obtained by the receiving end, the prediction is judged to be correct. If they are different, the prediction is judged to be wrong. Based on the judgment result, the initial prediction model is further trained with a new training set to obtain a trained prediction model and replace the initial prediction model in S1.

[0032] Beneficial effects: The present invention has the following advantages: 1. The vortex optical communication method and system described in the present invention have self-learning capabilities, can automatically mark data to construct training data sets, and reduce the need to manually mark large amounts of data; 2. It fully considers the impact of environmental factors such as receiving end position deviation, off-axis, camera bad pixels, and dirty mirror groups on the accuracy of information recognition. By continuously training the deep neural network model using pictures and labels carrying the latest environmental impacts, it continuously adaptively learns the impact of the current environment on the intensity distribution of the picture to ensure that the model has a high accuracy in judging the information carried in the picture; 3. It has anti-cracking function, and uses a dynamically changing code table and HMAC-SHA-1 encryption algorithm to encrypt the transmitted information to ensure the security of the information. At the same time, the deep neural network model continuously adapts to changes in the code table through adaptive learning, and can accurately identify encrypted information without directly obtaining a new code table. This strategy of combining dynamic encryption with deep learning effectively prevents information from being intercepted and cracked, and significantly enhances the security of information transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the process of this method;

[0034] Figure 2 This is the block diagram of the vortex optical communication system;

[0035] Figure 3 Schematic diagram of synthesis and encoding of holograms loaded on a digital micromirror array in a vortex optical communication system;

[0036] Figure 4 This is a diagram of the deep convolutional neural network structure;

[0037] Figure 5 A schematic diagram of data packet encoding rules;

[0038] Figure 6 This is a flowchart for deep convolutional neural network interpretation and training;

[0039] Figure 7 The focal field intensity patterns of the vortex beam captured by the camera under different environments;

[0040] Figure 8 Definition of code table B and code table B replacement diagram;

[0041] Figure 9 Schematic diagram of information transmission accuracy over a period of time. DETAILED DESCRIPTION

[0042] The technical solution of the present invention is described in detail below with reference to the embodiments and drawings.

[0043] The transmitting end of the present invention emits a vortex beam with different topological charges. Different topological charges can represent different information. The vortex beam is formed by parallel light irradiating the hologram loaded on the digital micromirror array. Figure 1 The 4-f system is obtained after filtering. The basic structure of the 4-f system is composed of two lenses with a focal length of f, which are coaxial and 2f apart. It has two functions: the first is to separate the spectrum of the light field and eliminate the zero-frequency components that do not carry information through filtering, thereby improving the signal-to-noise ratio of the information received by the subsequent system; the second is to cooperate with the digital micromirror array to enable the digital micromirror array that only has the ability to adjust the binary light field intensity to have the ability of complex field modulation. Therefore, in order to generate vortex light beams with different topological charges, in actual operation, it is only necessary to switch different holograms. At the receiving end, these vortex light beams can be converted into pictures with different intensity distributions, specifically through Figure 2 After the Daman grating and lens are focused onto a camera, the camera can capture images with different intensity distributions. These images with different intensity distributions correspond one-to-one to the vortex beams carrying topological charge. That is, each vortex beam emitted by the transmitter corresponds to an intensity distribution image at the receiver. Therefore, a deep neural network model deployed at the receiver can be used to identify the topological charge corresponding to the images with different intensity distributions, thereby obtaining the information sent by the transmitter.

[0044] Taking the eight permutations and combinations of the topological charge carried by the vortex beam as an example, namely -4, -3, -2, -1, 1, 2, 3, and 4, they correspond to an eight-bit binary number from high to low XXXXXXXX, and the binary number is the information sent by the transmitter.

[0045] For example, when the topological charge carried by the beam is -4, 4, the eight-bit binary number represented is 10000001; when the topological charge carried by the beam is -4, -2, 1, 3, the eight-bit binary number represented is 10101010; and when the topological charge carried by the beam is -4, -3, -2, -1, 1, 2, 3, 4, the eight-bit binary number represented is 11111111. For simplicity, the eight-bit binary number is represented in hexadecimal form, such as 11111111 as 0xFF, 00000000 as 0x00, and so on.

[0046] In communication systems, the basic unit of data transmission is the "packet." Scattered data must be packaged into packets before transmission. Otherwise, if some bits are lost or received incorrectly during transmission, all data must be retransmitted. Using packets as the basic structure, only the packets with errors need to be retransmitted. In the present invention, a packet includes header bits, data bits, flag bits, training data bits, and trailer bits.

[0047] If the receiving end does not receive the packet header during the process of receiving the data packet, it will continue to wait for the packet header. If it receives the packet header, it will start to receive the data bits. When at least four data bits are received, it will start to wait for the identification bit. If it does not receive the identification bit, it will continue to receive data. If it receives the identification bit, the data reception is completed, and it will start to receive training data and wait for the packet tail bit. If it does not receive the packet tail, it will continue to receive data. If it receives the packet tail, the entire data packet has been received.

[0048] This invention uses the transmission of a complete data packet as an example to explain how a receiving end receives information sent by a transmitting end. When code table B remains unchanged, the transmitting end generates a hologram and loads it onto a digital micromirror array. Parallel light irradiates the hologram, generating vortex light with a topological charge of "-4, -3, -2, -1, 1, 2, 3, 4." Upon reaching the receiving end, this vortex light is converted into an image by a Dammann grating and lens. This image is fed into a deep neural network model, which detects the eight bits of binary data it carries: "11111111," or 0xFF, representing the packet header. The receiving end then understands that the transmitting end will subsequently transmit real data (i.e., the corresponding data packet data bit type is transmission data). The transmitter generates vortex light based on the actual information to be transmitted (i.e., it converts the actual eight-bit binary information into ciphertext by querying code table A, and then generates vortex light encoded with this ciphertext). This vortex light reaches the receiver as an image, which is then fed into a deep neural network model for interpretation, where the specific binary information is output, completing the information transfer from the transmitter to the receiver. This is until a vortex light beam arrives from the transmitter. This beam carries the topological charge of "-4, -3, -2, -1, 1, 2, 3," which is converted by the receiver into an image and interpreted by the deep neural network model as the eight-bit binary data "11111110," or 0xFE, marking the end of the actual information transfer.

[0049] After the receiving end receives the intensity distribution image taken by the camera corresponding to the data bit, it uses the deep neural network model to interpret the eight-bit binary information contained in the image and outputs it. After receiving the identification bit, the receiving end enters the state of saving the image as a training image. That is, the receiving end will save the intensity distribution image taken by the camera and calculate an eight-bit binary information according to the TOTP rule. It will be used as a label to mark the intensity distribution image taken by the camera, forming a labeled training image and saving it for subsequent re-training of the deep neural network model.

[0050] To mitigate environmental influences on vortex beams (such as receiver position deviation, off-axis positioning, camera pixel defects, and lens contamination that reduce contrast, which can affect recognition accuracy), the deep neural network model must possess adaptive learning capabilities. This is achieved by labeling images captured by the receiving end with labels reflecting the latest environmental influences and feeding them into the neural network model for retraining. This allows the model to continuously learn the impact of the current environment on the image's intensity distribution, ensuring a consistently high level of accuracy in reading the eight-bit binary information carried in the image.

[0051] If code table B remains unchanged, the transmitter uses the HMAC-SHA-1 encryption algorithm and transforms the unchanged code table B to obtain the bytes 10101010. This generates a hologram with a beam topological charge of "-4, -2, 1, 3." This is then loaded onto the digital micromirror array, generating a vortex beam with topological charges of "-4, -2, 1, 3." The receiver then receives an intensity distribution image corresponding to the vortex beam, which carries information about the current environment.

[0052] Since the receiver shares the same private key and time system as the sender, using the HMAC-SHA-1 encryption algorithm (which encrypts the private key and Unix), it extracts a byte and obtains "10101010". The receiver then adds a byte label of "10101010" to the image. This image and byte label are stored in the dataset, thus completing the generation of labeled training data that carries the current environment information, thereby constructing a new dataset.

[0053] The sending end will also send this intensity distribution image to the currently trained deep neural network model to interpret the eight-bit binary information carried in the image. If the model's interpretation result is also "10101010", the judgment is recorded as correct; otherwise, it is recorded as incorrect. The accuracy rate of the model's judgment is calculated every 100 times. If the accuracy rate is less than 95%, a new dataset is used to further train the deep neural network model.

[0054] If the model's judgment accuracy remains greater than 95% during this period, and a total of 20,000 or more training images are collected in the dataset, the receiving end will use these 20,000 training images to further train the model, thereby replacing the old model to improve the model's tolerance to recent environmental changes, thereby achieving adaptive functionality.

[0055] To prevent a third party from obtaining code table B and copying the transmitted vortex beam, the present invention needs to have an anti-cracking function. That is, during the next data packet transmission, the transmitter immediately changes the code after transmitting the identification bit 0xFE and fixes it to the new code table B.

[0056] For example, if the sender obtains the byte 10101010 according to the HMAC-SHA-1 encryption algorithm, in the changed new code table B, "10101010" corresponds to "10100011", and the vortex beam topological charge corresponding to the byte 10100011 is "-4, -2, 3, 4", then it is loaded onto the digital micromirror array through the hologram to generate a vortex beam carrying the topological charge of "-4, -2, 3, 4".

[0057] The receiving end, unaware of code table B, decodes a single byte using the HMAC-SHA-1 encryption algorithm, resulting in "10101010." The receiving end then tags the received intensity distribution image with a byte label of "10101010." This intensity distribution image is then sent to a trained deep neural network model to interpret the eight bits of binary information contained within. Because the model was trained on a dataset with an unchanged code table, it interprets it as 10100011, which doesn't match the label received by the receiving end, resulting in an error. Since code table B has changed (generally speaking, after code table B is changed, a training set of at least 20,000 labeled images must be collected before code table B can be changed again), the model will make incorrect judgments in all subsequent judgments. If the model's accuracy rate per 100 judgments is less than 95%, the model will immediately enter a retraining phase. The receiving end will collect the intensity pattern distribution on the camera and label the pictures according to the aforementioned rules. It will collect at least 20,000 labeled pictures to form a training set to further train the model, thereby replacing the old model. The new model can implicitly learn the new code table B without directly obtaining the new code table B, thereby achieving the "anti-cracking" function.

[0058] Based on the above method, a third party cannot obtain a new code table B based only on the copied vortex beam. That is, the third party will decode the original code table B and the vortex beam with a topological charge of "-4, -2, 3, 4" to obtain "101000011" instead of the correct "10101010".

[0059] like Figure 1 As shown, based on the above content, the vortex optical communication method based on deep neural network described in the present invention, the sending end and the receiving end share the same private key and the same time system, including the following contents:

[0060] S1. Constructing a deep neural network initial model and training set, using the training set to train the deep neural network initial model to obtain an initial prediction model, including:

[0061] S101. Both the sending end and the receiving end obtain the current Unix timestamp, encrypt the private key and the Unix timestamp to generate a 20-digit ciphertext string, take the lower four bits of the ciphertext string as an offset, and obtain an eight-digit byte A starting from the offset.

[0062] For example, if the sender and receiver's private key string is "GI2TQMRVHAZDKOA" and use the same time system, and the current time is "2024-12-04 08:40:02 Beijing Time" and the Unix value is "1733301602," encrypting the private key and Unix value using the HMAC-SHA-1 algorithm yields the 20-bit ciphertext string "01111101101010011010." The lower four bits, "1010," are used as the offset. Since 1010 represents 10 in decimal, the offset is 10. Counting 10 bits from the right and taking an 8-bit byte, A, yields "11110110." Note that if the offset is 16, only 0111 is obtained, but this can be padded to a single byte, resulting in 00000111.

[0063] S102: At the sending end, the topological charge corresponding to byte A is obtained by querying code table B, and a vortex beam carrying the topological charge is generated based on the topological charge; at the receiving end, an intensity distribution image P corresponding to the vortex beam is obtained. a The receiving end uses the obtained byte A as the intensity distribution picture P a Byte tag.

[0064] Taking byte A as "11110110" as an example, the sender queries code table B based on the byte "11110110". At this time, if code table B has not changed, in code table B, "11110110" directly corresponds to "11110110", and the topological charge corresponding to "11110110" is "-4, -3, -2, -1, 2, 3". The generated hologram is loaded onto the digital micromirror array, generating a vortex beam carrying the topological charge number of "-4, -3, -2, -1, 2, 3".

[0065] The receiving end receives an intensity distribution image of a vortex beam carrying a topological charge of "-4, -3, -2, -1, 2, 3", marks the image with a byte tag of "11110110", and stores the image in the dataset after adding the tag.

[0066] The structure of the device that generates the vortex beam at the transmitting end and receives the intensity distribution image at the receiving end is as follows: Figure 2 As shown, it includes a light source module, a transmitting module and a receiving module. The light source module is used to generate parallel light with a suitable spot size and send the parallel light to the transmitting module.

[0067] The digital micromirror array in the transmitter module can be loaded with a series of coded holograms to quickly modulate parallel light into a vortex beam that can carry the transmitted information, such as Figure 3 As shown. The physical generation method of the coded hologram is:

[0068]

[0069] Where H cod is a fast time-varying coded hologram, where N is the number of bits that can be encoded by a single instantaneous vortex beam. A(n) is whether there is a topological charge in the beam. The binary representation of the component of , if it exists, it is recorded as 1, and if it does not exist, it is recorded as 0. The refresh rate of the time-varying coded hologram is the refresh rate of the digital micromirror array, represents the azimuth angle, and F(x,y) is the carrier frequency, so that the coded beam and the uncoded beam caused by the dead zone of the instrument are spatially separated on the focal plane of the 4-f system, so that only a series of coded vortex beams carrying different topological charges can be obtained.

[0070] The receiving module mainly consists of a focusing lens, a Dammann grating placed in front of the focusing lens, and a camera placed in the back of the focusing lens. The vortex light carrying information modulated by the transmitting module is incident in front of the Dammann grating. After the light intensity information is spatially separated by the Dammann grating, it is received by the camera and captured and saved as an image, such as Figure 7 The saved image is sent to the receiving module and first converted into an input data set that can be recognized by the deep neural network model. The deep neural network model then interprets it and directly outputs the information encoded in the vortex beam. From then on, the sending and receiving modules complete the information transmission.

[0071] S103, repeat this process until a certain number of intensity distribution images P with byte labels are obtained a An initial training set is constructed and a deep neural network model is trained based on the initial training set to enable it to learn the relationship between labels and intensity distribution images and obtain an initial prediction model.

[0072] like Figure 4 As mentioned above, it is the deep neural network model structure, and the loss function of the deep convolutional neural network is:

[0073]

[0074] Among them, y i,k is the true label of the i-th sample belonging to the k-th category. When sample i belongs to k, y i,k =1; otherwise y i,k =0. z i,k is the score logits of sample i for category k, N is the number of samples, and K is the number of categories.

[0075] S2. The sending end converts the actual information T to be sent into a ciphertext byte by querying the code table A. The sending end obtains the topological charge corresponding to the byte and generates the corresponding vortex beam based on the topological charge. At the receiving end, the intensity distribution image P corresponding to the vortex beam is obtained. t , using the initial prediction model to analyze the intensity distribution image P t Make an interpretation and output the intensity distribution picture P t The corresponding bytes are the information that the sender actually wants to send.

[0076] S3. After sending the ciphertext, the sending end and the receiving end obtain byte B through the method described in step 1 in S1. The sending end queries the code table B to obtain the topological charge corresponding to byte B, generates the corresponding vortex beam based on the topological charge, and obtains the intensity distribution image P corresponding to the vortex beam at the receiving end. b The receiving end will obtain byte B as the intensity distribution picture P b byte label, and the intensity distribution image P with byte B label b into the new training set.

[0077] S4. Using the initial prediction model to analyze the intensity distribution image P b The corresponding byte is predicted. If the predicted byte is the same as the byte B obtained by the receiving end, the prediction is judged to be correct. If they are different, the prediction is judged to be wrong, indicating that the prediction by the initial prediction model in S2 and the actual information to be sent obtained based on the prediction result are not completely correct. It is necessary to use a new training set to further train the initial prediction model to obtain a trained prediction model and replace the initial prediction model.

[0078] The reasons for the prediction errors include possible environmental interference in the intensity distribution image, which affects the prediction accuracy of the model; to resist cracking, code table B has been changed, that is, the correspondence between bytes and topological charges has changed, and the model has not learned the changed mapping relationship in time.

[0079] During the initial model training process, if the model learns the correct relationship between bytes and topological charges, any subsequent predictions of new bytes will be incorrect due to changes in code table B. If code table B changes, such as "10101010" corresponding to "10100011," and the byte "10100011" corresponds to a vortex beam with topological charges of "-4, -2, 3, 4," then the hologram is loaded onto the digital micromirror array, generating a vortex beam with topological charges of "-4, -2, 3, 4." However, the receiving end, unaware of code table B, will extract a byte and obtain "10101010" using the HMAC-SHA-1 encryption algorithm. The receiving end will then label the received intensity distribution image with the byte tag "10101010."

[0080] Although the model may be affected by changes in code table B or environmental factors, and the intensity distribution image corresponding to the same byte may change, the present invention uses the byte-selection rules of TOTP (Hash-based Message Authentication Code) and HMAC (Hash-based Message Authentication Code) to ensure that no matter what the intensity distribution image is, it will always be accurately labeled with byte labels. Therefore, this method regularly uses new data sets to adaptively train the deep neural network model to maintain the model's adaptability to the latest data distribution and maintain a high reading accuracy. The specific process is as follows:

[0081] S401, after a certain number of determinations, calculate the accuracy of the initial prediction model; if the accuracy is less than the set threshold, retrain the initial model using a new training set, and stop when the loss function converges;

[0082] S402, if the accuracy is greater than the set accuracy and the intensity distribution image P is obtained b If the number is greater than the set number, the initial prediction model will be retrained using the new training set, and the trained model will replace the initial prediction model in S1.

[0083] S5. During the information transmission process, repeat S2-S4 to improve the model's tolerance to recent environmental changes and achieve anti-cracking.

[0084] In the above process, when using the existing deep neural network model for image recognition to recover the information carried in the vortex light, the information is transmitted in the form of data packets as the basic unit, such as Figure 5 As shown in the figure, the transmission data bits in the data packet are directly output after being interpreted by the deep neural network model as the transmission information between the receiving end and the sending end; the training data bits will be saved as training images for further training of the deep neural network model. The training process is as follows Figure 6 shown.

[0085] The accuracy of the deep neural network model in recovering the information carried in the vortex light and the number of saved training images will be monitored in real time throughout the entire process of the system operation corresponding to this method. The current method for calculating the accuracy of the information recovery of the deep neural network model is:

[0086]

[0087] The total number of samples is usually M labeled training sets, and the total number of correctly classified samples is usually the number of samples whose actual values are the same as the labeled values after feeding these M labeled training sets into the current deep convolutional neural network model. M is a positive integer.

[0088] Furthermore, when the deep convolutional neural network model's accuracy in recovering information from vortex light falls below 95%, the system stops directly outputting that information and instead saves 20,000 training images. This retrains the deep convolutional neural network based on these images, relearning the perturbations and errors in the current system and attempting to improve the model's accuracy in recovering information from vortex light. Furthermore, when the number of saved training images exceeds 20,000, regardless of the current deep convolutional neural network model's accuracy in recovering information from vortex light, the system will retrain the deep convolutional neural network, thus completing its adaptive function.

[0089] During the communication process, the sending end can change the code table B at any time, such as Figure 8 As shown, the one-to-one correspondence between the binary information to be transmitted and the topological charge values carried by the vortex beam will change. For example, before code table B is changed, when the transmitter transmits the binary information "01110111", the topological charge values carried by the vortex beam are specifically "-3, -2, -1, 2, 3, 4". After code table B is changed, when the transmitter transmits the same binary information "01110111", the topological charge values carried by the vortex beam are specifically "-3, -2, -1". For the receiver, the changed code table B information is implicitly included in the training data bits in the data packet. By training the deep convolutional neural network with this labeled training data, the receiver can implicitly obtain the new code table B. However, since an eavesdropper cannot distinguish between data bits and training data bits, does not understand the encoding system centered on the HMAC algorithm and TOTP rules, and does not know the private key shared by the transmitter and receiver, it is difficult for them to obtain the new code table B before the transmitter changes code table B again. At this time, the eavesdropper sees that the information transmitted twice is different and has different practical meanings; while the recipient can implicitly obtain the new code table B by training a new deep convolutional neural network. Specifically, the recipient can still correctly parse the information to be transmitted as "01110111", thus realizing the anti-crackable vortex optical communication function.

[0090] like Figure 9 As shown, the information transmission accuracy of the present invention and the prior art over a period of time is demonstrated, indicating that the present invention continuously trains the deep neural network model by using pictures and labels carrying the latest environmental impact, so that it continuously adaptively learns the impact of the current environment on the image intensity distribution, which can ensure that the model has a high accuracy in interpreting the information carried in the picture.

Claims

1. A vortex optical communication method based on deep neural network, wherein the sending end and the receiving end share the same private key and the same time system, characterized in that: The following steps are involved: S1. Build an initial deep neural network model, train the initial model using the training set, and obtain an initial prediction model; S2: The sending end encrypts the information T through the code table A and generates a vortex beam to send. The receiving end receives the intensity distribution image P corresponding to the vortex beam. t , and use the initial prediction model to predict the intensity distribution image P t Make an interpretation and obtain information T; S3. After the information T is sent, the sender and receiver generate a ciphertext string based on the current Unix timestamp and private key, and obtain a byte B from the ciphertext string; the sender encrypts byte B using code table B and generates a vortex beam to send, and the receiver obtains the intensity distribution image P corresponding to the vortex beam. b , and use the obtained byte B as the intensity distribution picture P b byte labels, the intensity distribution image P b and labels are stored in a new training set; S4. Using the initial prediction model to analyze the intensity distribution image P b The corresponding byte is predicted. If the predicted byte is the same as the byte B obtained by the receiving end, the prediction is judged to be correct. If they are different, the prediction is judged to be wrong. Based on the judgment result, the initial prediction model is further trained with the new training set to obtain the trained prediction model, and the initial prediction model in S1 is replaced; S5. Repeat S2-S4 to achieve secure information transmission between the sending and receiving ends.

2. The vortex optical communication method based on deep neural network according to claim 1, characterized in that The method for obtaining the initial prediction model described in S1 is: S101. Both the sending and receiving ends obtain the current Unix timestamp, encrypt the private key and the Unix timestamp to generate a l-bit ciphertext string, take the low n bits or high n bits of the ciphertext string as an offset, and obtain an m-bit byte A starting from the offset. <l; S102: The transmitting end encrypts the acquired byte A by querying the code table B and generates a vortex beam to be sent; the receiving end obtains the intensity distribution image P corresponding to the vortex beam after passing through the Dammann grating and focusing lens. a The receiving end uses the obtained byte A as the intensity distribution picture P a Byte tag of S103. Repeat steps S101-S112 until a certain number of intensity distribution images with byte labels are obtained to construct an initial training set, train the deep neural network model based on the initial training set, learn the relationship between the labels and the intensity distribution images, and obtain an initial prediction model.

3. The vortex optical communication method based on deep neural network according to claim 2, characterized in that The code table A includes a mapping relationship between information and bytes, and the code table B includes a mapping relationship between bytes and topological charges.

4. The vortex optical communication method based on deep neural network according to claim 2, characterized in that The transmitting end generates a vortex beam carrying a topological charge based on the topological charge, and the receiving end obtains the intensity distribution image corresponding to the vortex beam by generating a hologram about a specific topological charge and loading it onto a digital micromirror array. Parallel light illuminates the digital micromirror array to generate a vortex beam carrying the topological charge. The vortex beam passes through a Dammann grating and a lens and is focused onto a camera, thereby obtaining an intensity distribution image.

5. The vortex optical communication method based on deep neural network according to claim 1, characterized in that In S2, the sending end encrypts the information T and sends it. The receiving end decrypts the information T after receiving it. The sending end converts the actual information T to be sent into bytes in ciphertext by querying the code table A. The sending end converts the bytes in ciphertext into the corresponding topological charge number, generates the corresponding vortex beam based on the topological charge number, and obtains the intensity distribution image P corresponding to the vortex beam at the receiving end. t , using the initial prediction model to analyze the intensity distribution image P t Make an interpretation and output the intensity distribution picture P t The corresponding bytes are the information that the sender actually wants to send.

6. The vortex optical communication method based on deep neural network according to claim 1, characterized in that In S4, the initial prediction model is further trained using the new training set based on the judgment result. The method for obtaining the trained prediction model is as follows: S401, after a certain number of determinations, calculate the accuracy of the initial prediction model; if the accuracy is less than the set threshold, retrain the initial model using a new training set, and stop when the loss function converges; S402: If the accuracy is greater than the set accuracy and the number of intensity distribution images stored in the new training set is greater than the set number, the initial prediction model is retrained using the new training set, and the trained model is used to replace the initial prediction model in S1.

7. The vortex optical communication method based on deep neural network according to claim 1, characterized in that The information transmission between the sending end and the receiving end is based on the data packet as the basic unit, including the packet header bit, data bit, identification bit, training data bit, and packet tail bit. If the receiving end does not receive the packet header during the process of receiving the data packet, it will continue to wait for the packet header. If it receives the packet header, it will start receiving the data bits. When at least four data bits are received, it will start waiting for the identification bit. If it does not receive the identification bit, it will continue to receive data. If it receives the identification bit, the data reception is completed, and it will start receiving training data and wait for the packet tail bit. If it does not receive the packet tail, it will continue to receive data. If it receives the packet tail, the entire data packet is received.

8. The vortex optical communication method based on deep neural network according to claim 7, characterized in that: The data bits are the ciphertext of the information to be transmitted, and the actual information can be directly output after being interpreted by the neural network at the receiving end.

9. The vortex optical communication method based on deep neural network according to claim 7, characterized in that: The training data bits following the identification bits are used to collect training data.

10. A vortex optical communication system based on a deep neural network, wherein the transmitting end and the receiving end share the same private key and the same time system, characterized in that: Includes the following modules: Initial prediction model generation module: builds the initial model of the deep neural network, trains the initial model using the training set, and obtains the initial prediction model; Information transmission module: The sending end encrypts the information T through the code table A and generates a vortex beam to send, and the receiving end receives the intensity distribution image P corresponding to the vortex beam t , and use the initial prediction model to predict the intensity distribution image P t Make an interpretation and obtain information T; Training set generation module: After the information T is sent, the sender and receiver generate a ciphertext string based on the current Unix timestamp and private key, and obtain a byte B from the ciphertext string; the sender encrypts byte B through code table B and generates a vortex beam to send, and the receiver obtains the intensity distribution image P corresponding to the vortex beam b , and use the obtained byte B as the intensity distribution picture P b byte labels, the intensity distribution image P b and labels are stored in a new training set; Initial prediction model training module: Use the initial prediction model to train the intensity distribution image P b The corresponding byte is predicted. If the predicted byte is the same as the byte B obtained by the receiving end, the prediction is judged to be correct. If they are different, the prediction is judged to be wrong. Based on the judgment result, the initial prediction model is further trained with a new training set to obtain a trained prediction model and replace the initial prediction model in S1.

Images