Spectrum encryption method based on diffraction deep neural network

Abstract

The invention discloses a spectrum encryption method based on a diffraction deep neural network, and the method comprises the steps: training a diffraction deep neural network model, and building a spectrum password book, i.e., enabling the diffraction deep neural network to map each input wavelength to an output pattern, i.e., a spectrum basis, under the condition of spatial coherence or spatial incoherence; generating a spectrum password of the encrypted information, namely generating a corresponding spectrum password according to the encrypted information and the constructed spectrum password book; and decrypting the information by using the spectrum password book and the spectrum password, namely inputting a plane wave corresponding to the spectrum amplitude according to the positive and negative spectrum passwords, synthesizing corresponding positive and negative encrypted information through a diffraction deep neural network, and decrypting the encrypted information. The spectrum encryption complexity is low, and the security level is high.

Description

Technical Field

[0001] The present invention relates to the technical field of information encryption, and in particular to a spectral encryption method based on a diffraction deep neural network. Background Art

[0002] Optical neural networks (PNNs) use the properties of light for computing, providing speed and efficiency that surpass traditional electronic neural networks. They use the parallel processing capabilities and high bandwidth of optical systems to significantly reduce the energy consumption and latency of tasks such as data processing, machine learning, and real-time analysis. Among various optical neural networks, diffractive deep neural networks (D2NNs) have received continued and significant attention due to their simplicity, high efficiency, strong generalization ability, and scalability. D2NNs consist of multiple modulation layers that adjust the amplitude or phase of the light field and are optimized through deep learning methods to achieve arbitrary input-to-output mapping functions. This approach solves various artificial intelligence challenges at the speed of light and provides an innovative computing solution for the post-Moore's Law era.

[0003] Information encryption is essential for protecting information security and has been widely used in fields such as telecommunications, data storage, and financial transactions. Traditional electronic encryption methods face many problems, including complex algorithm design, limited security levels, long delays, and high computing power requirements. In contrast, optical encryption offers significant advantages such as higher speed, lower energy consumption, parallel processing capabilities, and numerous adjustable optical dimensions, effectively addressing the limitations of electronic encryption. Encryption using optical neural networks further enhances these advantages, providing more flexible design options and intelligent control by exploiting the intrinsic properties of light. Summary of the invention

[0004] In order to solve the problems of complex algorithm design, limited security level, long delay and high computing power requirements in traditional electrical encryption technology, the present invention provides a spectral encryption method for information encryption through a diffraction deep neural network, which uses the spectral dimension of light for optical encryption. A spectral codebook is constructed to map each wavelength to a specific, self-designed output pattern under spatial coherence and incoherence conditions. Subsequently, the spectral password corresponding to any encrypted information in a specific spectral codebook is calculated, thereby perfectly encrypting the information into a spectral codebook and a spectral password. The present invention provides a reference for the application of diffraction deep neural networks in spectral interaction.

[0005] The present invention discloses a spectrum encryption method based on a diffraction deep neural network, which comprises:

[0006] Training the diffraction deep neural network model and building a spectral codebook, that is, the diffraction deep neural network maps each input wavelength to an output pattern, i.e., the spectral basis, under spatial coherence or spatial incoherence conditions;

[0007] Generate a spectral password for the encrypted information, that is, generate a corresponding spectral password according to the encrypted information and the constructed spectral password book;

[0008] The spectral code book and spectral password are used to decrypt information. That is, according to the positive and negative spectral passwords, the plane wave with the corresponding wavelength amplitude is input, and the corresponding positive and negative encrypted information is synthesized through the diffraction deep neural network to decrypt the encrypted information.

[0009] Furthermore, the input wavelength selection of the spectral codebook, that is, the number of input working wavelengths and the resolution of the working wavelengths are selected; for the encrypted information It requires n input wavelengths, with arbitrary intervals between them;

[0010] The output pattern design of the spectral codebook, that is, the output pattern corresponding to each wavelength is designed as the spectral basis; Design n spectral bases, which are linearly independent from each other, to form a maximal linearly independent group of rank n

[0011] Furthermore, the input of the diffraction deep neural network is a unit amplitude surface light source with spatial coherence or incoherence, and the unit amplitude surface light source includes multiple spectral bands λ i ,i=1,…,n; the amplitude of the light source in each spectral channel is uniform and constant; for spatially coherent input, the phase is a real constant, while for spatially incoherent input, the phase varies randomly at different spatial positions.

[0012] Furthermore, in the kth propagation process, the input light field of the surface light source with different wavelengths after passing through the rectangular aperture is expressed as:

[0013]

[0014] in, represents the input light field, rect represents the rectangular aperture function, represents the phase at the spatial position (x, y) during the kth propagation process; λ i represents the i-th wavelength channel;

[0015] For each wavelength channel λ i , the unit amplitude is evenly distributed on the entire plane and remains constant over time. After the diffraction deep neural network transformation, the light field distribution on the output plane is given by:

[0016]

[0017] in, represents the output light field intensity distribution of the i-th channel during the k-th propagation process, Represents the modulation effect of the diffractive deep neural network, using pure phase modulation.

[0018] Furthermore, the spectral band λ i The output intensity of is the spectral basis of the spectral band, expressed as:

[0019]

[0020] Among them, N tr represents the number of repetitions of generating random phases during the training of the diffraction deep neural network model, λ i is the i-th spectral band.

[0021] Furthermore, two types of loss functions are used to collaboratively train the diffractive deep neural network model:

[0022] L=L dis +β·L balance

[0023] Among them, L represents the total loss function, β is the distribution loss function L dis And the balance loss function L balance The weight coefficient, L dis The goal is to minimize the mean square error between each spectral basis and its corresponding true value:

[0024]

[0025] in, It is a spectral basis The corresponding true value, is the maximum intensity value among all pixels on the output plane, and HW represents the total number of pixels in each spectral basis;

[0026] L balance The goal is to maximize the intensity of each spectral basis As close as possible:

[0027]

[0028] After the diffraction deep neural network model training is completed, the diffraction deep neural network is used as the spectral codebook.

[0029] Furthermore, after constructing the spectral codebook, each wavelength corresponds to a non-negative real-valued spectral basis Any vector It is expressed as:

[0030]

[0031] Among them, α i Representation and spectral basis vector The relevant weights;

[0032] Constructing a matrix Get the spectral basis corresponding to each wavelength as a column: Then I is expressed as:

[0033]

[0034] Among them, α=[α1,α2,…,α n ] T , Indicates the λth i spectral basis;

[0035] matrix is reversible, then There is a unique solution, namely

[0036] The weight vector α is decomposed into the positive weight vector and negative weight vector Among them, α + and α - The component of is defined as:

[0037]

[0038] I is decomposed into two components: I + and I - , I + Corresponding to the positive weight vector α + , I - Corresponding to the negative weight vector α - ;

[0039] I + and I - are all non-negative and are formed by non-negative linear combinations of non-negative spectral bases. The spectral bands are simultaneously input on the input plane of the diffraction deep neural network, and the required output intensity I is synthesized on the output plane. + and I - ;

[0040] The light intensity I obtained by double optical path or two measurements + and I - Obtaining encrypted information I;

[0041] Positive Branch I + and negative branch I - The corresponding spectral amplitude sequences, namely the positive spectral code and the negative spectral code, together constitute the spectral code.

[0042] Further, I is expressed as:

[0043]

[0044] Furthermore, the output intensity I + and I - Respectively expressed as:

[0045]

[0046]

[0047] Further, the positive spectrum code and the negative spectrum code are expressed as: and

[0048] Due to the adoption of the above technical solution, the present invention has the following advantages:

[0049] 1. The present invention adopts a diffraction-based deep neural network and uses the spectral dimension of light to convert information into a spectral codebook and spectral key to achieve intelligent and high-level encryption of information.

[0050] 2. The present invention uses a diffraction deep neural network to decrypt information, achieving near-light-speed, low-energy consumption, and high-quality information decryption.

[0051] 3. The method proposed in the present invention has strong robustness and is applicable to more complex environments; it has strong flexibility and the codebook construction is flexible and changeable; it has strong scalability and can be further combined with the electrical module to achieve higher-level encryption or a simpler architecture. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0053] Figure 1 A schematic diagram of a process of a spectral encryption method based on a diffraction deep neural network according to an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of a spectral codebook construction process according to an embodiment of the present invention;

[0055] Figure 3 Schematic diagram of the spectral key generation process and information reconstruction process according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present invention is further described in conjunction with the accompanying drawings and embodiments, and the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present invention.

[0057] See also Figure 1 The present invention provides an embodiment of a spectrum encryption method based on a diffraction deep neural network, which comprises the following steps:

[0058] S1, training the diffraction neural network model and building the spectral codebook:

[0059] In this embodiment, the input and output of the spectral codebook include input wavelength selection and output pattern design;

[0060] Input wavelength selection, that is, select the number of input working wavelengths and the resolution of the working wavelengths. It requires n input wavelengths, with arbitrary intervals between wavelengths; output pattern design, that is, designing the output pattern corresponding to each wavelength as the spectral basis. It is necessary to design n spectral bases, which are linearly independent from each other, to form a maximal linearly independent group of rank n.

[0061] In this embodiment, the training diffraction neural network model in S1 builds a spectral codebook, that is, each input wavelength is mapped to an output pattern (spectral basis) based on diffraction deep neural networks (D2NNs) under spatial coherence or spatial incoherence conditions;

[0062] like Figure 2 As shown, the input is a unit amplitude surface light source with spatial coherence or incoherence, containing multiple spectral bands (λ i ,i=1,…,n). In the entire plane, the amplitude of the light source of each spectral channel is uniform and constant. For spatially coherent input, the phase is a real constant, while for spatially incoherent input, the phase varies randomly at different spatial positions. In the kth propagation process, the input light field after passing through the rectangular aperture is expressed as:

[0063]

[0064] in, represents the phase at the spatial position (x, y) during the kth propagation process. For each wavelength channel λ i , the unit amplitude is evenly distributed over the entire plane and remains constant over time. 2 After NN transformation, the light field distribution on the output plane is given by:

[0065]

[0066] in, represents the output light field intensity distribution of the i-th channel during the k-th propagation process, Indicates D 2 The modulation effect of NN adopts pure phase modulation. For spatially coherent samples, the output remains constant over time. However, for spatially incoherent samples, the phase changes over time, causing the output to change accordingly. During the training process, in order to simulate the spatial incoherence effect, the output intensity distribution is the average of the output intensity distribution during multiple forward propagation processes calculated according to the intensity superposition principle. Since the modulation layer remains constant over time, the input in these forward propagations passes through the same modulation function. Let N ir represents the number of repetitions of generating random phases during training, λ i The output intensity of a spectral band, i.e. the spectral basis of the band, can be expressed as:

[0067]

[0068] To facilitate subsequent explanations, when considered as a vector, λ i Output light intensity of spectral band Expressed as

[0069] In order to achieve the specified, self-designed spectral basis while balancing the intensity differences between the spectral basis, two types of loss functions are used for joint training:

[0070] L=L dis +β·L balance

[0071] Among them, L represents the total loss function, β is the distribution loss function L dis And the balance loss function L balance The weight coefficient of L is set to β = 0.1. dis The goal is to minimize the mean square error between each spectral basis and its corresponding true value:

[0072]

[0073] in, is the corresponding true value, L refers to the maximum intensity value among all pixels on the output plane, and HW represents the total number of pixels in each spectral basis. balance The goal is to maximize the intensity of each spectral basis As close as possible:

[0074]

[0075] After training, D 2The NN acts as a spectral codebook, mapping each input spectral band to a unique output, constructing a spectral codebook containing n linearly independent vectors of n dimensions.

[0076] S2, generate the spectral password of the encrypted information:

[0077] In this embodiment, the spectrum password for generating the encrypted information in S2 is based on the encrypted information With the constructed spectral codebook, the corresponding spectral password is generated.

[0078] After constructing the spectral codebook, each wavelength corresponds to a non-negative real-valued spectral basis These bases are linearly independent and together open up a real space For the convenience of explanation, when the spectral basis is used as the output light intensity of the output surface, it is not bolded, that is, it is expressed as When the spectral basis is used as a vector, it is expressed in bold, that is, I,I + and I - Similarly;

[0079] Any vector can be represented by these baseline properties:

[0080]

[0081] Among them, α i Representation and spectral basis The relevant weights. Construct a matrix Get its spectral basis as columns: Then I can be expressed as:

[0082]

[0083] Among them, α=[α1,α2,…,α n ] T Since these spectral bases are linearly independent, the matrix is reversible, then There is a unique solution, namely Therefore, the weight vector α can be solved directly without obtaining it through training.

[0084] The weight vector α can be decomposed into the positive weight vector and negative weight vector where α + and α - The components of are defined as follows:

[0085]

[0086] Similarly, I can also be decomposed into two components: I+ , corresponding to the positive weight vector α + , and I - , corresponding to the negative weight vector α - . Therefore, I can also be expressed as:

[0087]

[0088] like Figure 3 As shown, due to I + and I - are all non-negative and are formed by non-negative linear combinations of non-negative spectral bases, so they can be 2 Spectral bands with specific amplitudes are simultaneously input to the input plane of the NN, thereby synthesizing the required output intensity I on the output plane. + and I - :

[0089]

[0090]

[0091] The light intensity I obtained by double optical path or two measurements + and I - By performing simple electronic subtraction, the encrypted information I can be obtained. The spectral amplitude sequences corresponding to the positive branch and the negative branch, namely the positive spectral password and the negative spectral password, can be expressed as: and Together, they constitute the spectral password. Therefore, the spectral encryption of information is realized using the diffractive deep neural network, in which the information is hidden as the spectral codebook and the spectral password. Only by obtaining the spectral codebook and the spectral password at the same time can the encrypted information be reconstructed.

[0092] S3, decrypt information using the spectral codebook and spectral password:

[0093] In this embodiment, S3 uses the spectral code book and spectral password to decrypt information, that is, according to the positive and negative spectral passwords, a plane wave with corresponding wavelength amplitude is input, and the corresponding positive and negative encrypted information is synthesized through the diffraction deep neural network, so as to decrypt the encrypted information by subtraction.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A spectrum encryption method based on diffraction deep neural network, characterized in that: include: Training the diffraction deep neural network model and building a spectral codebook, that is, the diffraction deep neural network maps each input wavelength to an output pattern, i.e., the spectral basis, under spatial coherence or spatial incoherence conditions; Generate a spectral password for the encrypted information, that is, generate a corresponding spectral password according to the encrypted information and the constructed spectral password book; The spectral code book and spectral password are used to decrypt information. That is, according to the positive and negative spectral passwords, the plane wave with the corresponding wavelength amplitude is input, and the corresponding positive and negative encrypted information is synthesized through the diffraction deep neural network to decrypt the encrypted information.

2. The method according to claim 1, characterized in that: The input wavelength selection of the spectral codebook, i.e., the number of input working wavelengths and the resolution of the working wavelengths are selected; for the encrypted information It requires n input wavelengths, with arbitrary intervals between them; The output pattern design of the spectral codebook, that is, the output pattern corresponding to each wavelength is designed as the spectral basis; Design n spectral bases, which are linearly independent from each other, to form a maximal linearly independent group of rank n 3. The method according to claim 1, characterized in that The input of the diffraction deep neural network is a unit amplitude surface light source with spatial coherence or incoherence, and the unit amplitude surface light source contains multiple spectral bands λ i ,i=1,…,n; the amplitude of the light source in each spectral channel is uniform and constant; for spatially coherent input, the phase is a real constant, while for spatially incoherent input, the phase varies randomly at different spatial positions.

4. The method according to claim 3, characterized in that In the kth propagation process, the input light field after the surface light sources of different wavelengths pass through the rectangular aperture is expressed as: in, represents the input light field, rect represents the rectangular aperture function, represents the phase at the spatial position (x, y) during the kth propagation process; λ i represents the i-th wavelength channel; For each wavelength channel λ i , the unit amplitude is evenly distributed on the entire plane and remains constant over time. After the diffraction deep neural network transformation, the light field distribution on the output plane is given by: in, represents the output light field intensity distribution of the i-th channel during the k-th propagation process, Represents the modulation effect of the diffractive deep neural network, using pure phase modulation.

5. The method according to claim 4, characterized in that Spectral band λ i The output intensity of is the spectral basis of the spectral band, expressed as: Among them, N tr represents the number of repetitions of generating random phases during the training of the diffraction deep neural network model, λ i is the i-th spectral band.

6. The method according to claim 4 or 5, characterized in that: Two types of loss functions are used to jointly train the diffractive deep neural network model: L=L dis +β·L balance Among them, L represents the total loss function, β is the distribution loss function L dis And the balance loss function L balance The weight coefficient, L dis The goal is to minimize the mean square error between each spectral basis and its corresponding true value: in, It is a spectral basis The corresponding true value, is the maximum intensity value among all pixels on the output plane, and HW represents the total number of pixels in each spectral basis; L balance The goal is to maximize the intensity of each spectral basis As close as possible: After the diffraction deep neural network model training is completed, the diffraction deep neural network is used as the spectral codebook.

7. The method according to claim 6, characterized in that After constructing the spectral codebook, each wavelength corresponds to a non-negative real-valued spectral basis Any vector It is expressed as: Among them, α i Representation and spectral basis The relevant weights; Construct a matrix Use the spectral basis corresponding to each wavelength as the column vector of the matrix: Then I is expressed as: Where, α=[α1,α2,…,α n ] T ; matrix is reversible, then There is a unique solution, namely The weight vector α is decomposed into the positive weight vector and negative weight vector In, α + and α - The component of is defined as: I is decomposed into two components: I + and I - , I + Corresponding to the positive weight vector α + , I - Corresponding to the negative weight vector α - ; I + and I - are all non-negative and are formed by non-negative linear combinations of non-negative spectral bases. The spectral bands are simultaneously input on the input plane of the diffraction deep neural network, and the required output intensity I is synthesized on the output plane. + and I - ; The light intensity I obtained by double optical path or two measurements + and I - Obtaining encrypted information I; Positive branch I + and negative branch I - The corresponding spectral amplitude sequences, namely the positive spectral code and the negative spectral code, together constitute the spectral code.

8. The method according to claim 7, characterized in that I is represented by:

9. The method according to claim 7 or 8, characterized in that: Output intensity I + and I - Respectively expressed as:

10. The method according to claim 7, characterized in that That is, the positive spectrum code and the negative spectrum code are expressed as: and

Images