A carrier-free encrypted transmission method based on a 3D model

Abstract

This invention discloses a carrier-free binary sequence transmission method and apparatus based on voxelization and geometric transformation. The sender encrypts the binary secret information to be transmitted (batch number, production date, and signature combination) and combines it with parameters such as voxelization threshold, resolution, slice axis, and modulus to form a text index. This index, along with a pre-set 3D model library, is input into the encoding neural network, outputting a triplet consisting of a model identifier, a 3D rotation angle, and voxelization parameters. Only this triplet is transmitted in the public channel. The receiver, based on the same model library, loads a model according to the model identifier, performs geometric decoding according to agreed-upon geometric transformation and voxelization rules to obtain the binary sequence, and decrypts it using a pre-shared key to recover the secret information. This invention achieves 3D printing anti-counterfeiting verification and authorization control by comparing the decrypted information, featuring no explicit carrier transmission, high concealment, and strong security.

Description

Technical Field

[0001] This invention relates to the field of information security technology, and in particular to a carrier-free information encryption transmission method based on a three-dimensional model. Background Technology

[0002] The goal of steganography is to transmit information without attracting attention. Current mainstream methods mostly use digital carriers such as two-dimensional images and audio as hosts, carrying ciphertext by making minor modifications to the carrier's pixels or sampling points. Because these modifications alter the statistical characteristics and distribution of the carrier, they are easily detected by modern steganalysis models and are prone to loss or distortion after multiple compression, resampling, and filtering processes. Furthermore, the "carrier-free" image steganography that has emerged in recent years typically uses generative models to directly synthesize ciphertext-carrying images; however, the generation quality, controllability, and reproducibility are greatly affected by the training data and model stability, making stable implementation in engineering difficult.

[0003] With the widespread availability of 3D data, some literature has attempted to embed secret information into 3D meshes or point clouds, for example, through vertex quantization, normal perturbation, or mesh rearrangement. However, these methods still require geometric modifications to the 3D model itself, making them susceptible to corruption during common 3D processing steps such as simplification, reconstruction, resampling, smoothing, and rotation / scaling normalization. Furthermore, the modified 3D files need to be transmitted over public channels, resulting in large file sizes, high transmission costs, and increased risks of forensics and comparison. The approach of directly "synthesizing 3D models to carry ciphertext" based on generative models also faces issues of insufficient geometric fidelity, topological consistency, and repeatability, making it difficult to meet the stability and verifiability requirements of secure communication scenarios. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a carrier-free information encryption transmission method based on a three-dimensional model. Using geometric transformation as the public transmission parameter, and combining the voxelization statistics of the three-dimensional model to obtain a unique and verifiable binary sequence, it is possible to complete the encoding and decoding of secret information without changing the three-dimensional model itself or transmitting a carrier. Technical solution

[0005] Compared with the prior art, the present invention has the following significant advantages.

[0006] 1. This invention entrusts the inverse mapping from the target bit sequence to the triplet, which is difficult to solve manually, to be automatically learned by the neural network at the encoding end. It extracts voxel geometric features through three-dimensional convolution and index text features through text encoder. In the multimodal Transformer backbone, it adopts a cross-attention and convolutional multilayer perceptron (CLMP) structure to fuse the two. Only on the brand server side, the three-axis rotation angle that meets the constraints is regressed. Thus, while ensuring the simple and controllable structure of the decoding end, it significantly reduces the search complexity of the encoding end and expands the available angle space and bit capacity.

[0007] 2. The decoding end of this invention adopts deterministic geometric and voxel statistical rules to normalize, voxelize, and count slices of the shared model according to the received rotation angle and perform modulo operations to recover the binary sequence. The decoding process does not rely on neural network parameters, making it suitable for embedding into IoT (Internet of Things) devices with limited computing power and storage, such as 3D printing control boards and industrial gateways, which facilitates engineering implementation and security auditing.

[0008] 3. The present invention employs nonlinear encryption and modulus extraction, which complicates the relationship between the original geometric features and the final code, increases the difficulty of inferring the precise geometric shape from the code sequence, and improves the encryption effect.

[0009] 4. This invention allows for a flexible trade-off between information capacity and robustness by adjusting parameters such as voxel resolution G, slice axis, and modulus m. For example, when the modulus m=2, each voxel slice carries 1 bit of information, exhibiting high anti-interference capability; when the modulus m=4, each slice can encode 2 bits of information, providing greater data carrying capacity within the same model and rotation range. Combined with business requirements, this invention can achieve secure transmission of tens to hundreds of bits of information in scenarios such as offline 3D printing authorization, anti-counterfeiting verification of parts, and secure device pairing.

[0010] 5. This invention employs anti-aliasing occupancy estimation, which makes the occupancy value of voxels no longer a simple 0 or 1, but a "soft occupancy" that can take intermediate values. When the model undergoes a small shift, the change in voxel values ​​is smooth and continuous, rather than abrupt. This smoothness ensures that the subsequent column statistics can also change stably, making the final generated digital sequence insensitive to small geometric perturbations, thereby significantly improving the robustness and anti-interference ability of the entire encoding scheme. Attached Figure Description

[0011] Figure 1 This is the overall flowchart of the present invention.

[0012] Figure 2 This is a schematic diagram of the neural network encoder of the present invention.

[0013] Figure 3 This is a schematic diagram of the voxelization process. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0015] In many practical scenarios, communicating parties can share a publicly available library of 3D models, which provides the conditions for "parameter-transmitted" carrier-free information hiding: by transmitting only a small number of parameters over a public channel, the receiving end can recover the ciphertext using the same 3D model and agreed-upon rules. This invention, targeting shared 3D models, proposes a deterministic mapping mechanism based on geometric transformations and voxelization statistics, which can stably map and recover the secret bit sequence without modifying the 3D model itself or transmitting the carrier.

[0016] The main solution of this invention is as follows: at the sending end, in response to the information encryption transmission command, an index text containing binary secret information and voxelization configuration parameters is constructed; a pre-trained coding neural network is used to fuse the index text with the geometric features of the candidate 3D model to regress the model identifier and 3D rotation angle of the target 3D model; at the receiving end, based on the model identifier and 3D rotation angle, the binary secret information is recovered through deterministic geometric transformation and voxelization statistical rules.

[0017] like Figure 1 As shown, the specific implementation steps of the carrierless information encryption transmission method based on a three-dimensional model of the present invention are as follows.

[0018] Step 1: The sender first generates binary ciphertext based on the binary secret information to be transmitted (a combination of product batch number, production date and manufacturer's signature), and then combines the binary ciphertext with its corresponding voxelization configuration parameters such as voxelization threshold, voxel resolution, slice axis, and modulus number to form a text index.

[0019] Step 2: Input the text index and the geometric models in the preset 3D model library into the encoding neural network. The network outputs the model identifier, 3D rotation angle and voxelization parameter triplet of the target 3D model.

[0020] Specifically, such as Figure 2 The diagram shows the neural network encoder of the present invention. In this embodiment, the terminal device to which the information encryption transmission device belongs includes at least an output module, a processor, a memory, and a communication module. The memory stores an information encryption transmission program, which, when executed by the processor, implements the steps of the method described in this invention.

[0021] This invention constructs a neural network encoder, which aims to achieve the following mapping: given a target binary sequence, it regresses a 3D model selected from a shared 3D model library and regresses the 3D rotation angle and voxelization parameters of that model. The overall structure of the neural network encoder mainly consists of the following modules: The 3D geometric feature encoding module normalizes each 3D point cloud model in the shared model library and performs voxelization according to resolution to obtain the occupancy tensor. .Will The input is fed into a 3D convolutional encoder, and after passing through multiple layers of 3D convolution, normalization, and nonlinear activation GELU, a three-dimensional feature map is obtained. By using 1×1 convolution to transform the number of channels into a uniform feature dimension, and dividing the 3D feature volume into local sub-blocks and flattening them, a geometric token sequence is obtained. And add global aggregation to the front end. These tokens can be pre-computed offline and cached as a "geometric dictionary".

[0022] For each 3D point cloud model in the shared dataset Normalization is performed, and the tensor is scaled and translated to fit within the unit cube. Then, it is voxelized to voxel resolution to obtain the occupancy tensor. The voxel resolution G can be set within the range of 4, 8, 16, 32, and 64 depending on storage and computational resources to balance geometric detail preservation with computational overhead.

[0023] voxel tensors The input is fed into a 3D convolutional encoder, and after multiple layers of 3D convolution, normalization, and non-linear activation GELU, a 3D feature map is obtained:

[0024] It is the original number of feature channels. This indicates that, while maintaining the original spatial dimensions, the number of channels is transformed into a uniform feature dimension C using 1×1 convolution, resulting in the feature volume.

[0025] The 3D feature volume is then divided into several local sub-blocks according to rules and flattened to obtain the geometric token sequence of the model:

[0026] Add an extra special token for global aggregation to the beginning of the sequence, denoted as... This is used to store comprehensive information at the model level. Since the model library is a finite set... The aforementioned geometric tokens can be pre-calculated and cached offline at the encoding end. During actual encoding, the corresponding token sequence can be indexed from the dictionary based on the model index (model id) selected by the network.

[0027] Where is the number of candidate 3D models, and k is the index for iterating through each 3D model.

[0028] The binary sequence and voxel data feature encoding modulus concatenates the target binary sequence and business-related fields with candidate voxelization parameters (resolution, slice axis, modulus) to form an index text. For example, bits = 1010..; len = 128; voxel = 128; axis = z; mod = 4. This index text is input into the CLIP text encoder, and after tokenization, embedding, and linear mapping, a text token sequence of dimension is obtained. .

[0029] The encoding end needs to reconstruct a set of triplets containing a 3D model identifier, rotation angle, and voxelization parameters, given a target bit sequence, to recover the required bit sequence at the decoding end. Therefore, this invention organizes the target binary sequence and related parameters into structured text, which serves as the second modal input.

[0030] Specifically, for a single encoding task, the target bit sequence and business-related fields are first given, along with candidate voxelization parameters (including voxel resolution, slice axis, and modulus). This information is then concatenated into an index text containing L printer process parameters, the compressed encoding of the printer ID, the authorized machine number, delivery date, print count limit, and the number of message authentication codes, among other data parameters.

[0031] For each 3D model in the database, the rotation step size for each dimension of the corresponding 2 to the power of n voxel parameter is set to 5 degrees to generate a text dataset.

[0032] The indexed text input CLIP binary sequence and voxelized configuration feature encoding module, after word segmentation and embedding, yield a text token sequence. The text token dimension is transformed from the original text encoding dimension through a linear mapping. Mapped to the same dimensions as the geometry token, we get:

[0033] in, Let represent the j-th text feature token after dimension mapping. The number of text tokens. The text token, bit token, and geometry token of each candidate model are input together into the subsequent multimodal Transformer model to achieve joint modeling that "selects the appropriate model and geometric parameters starting from the bit".

[0034] The multimodal Transformer and CLMP fusion module concatenates the geometric token sequence and the text token sequence, and inputs them into the multi-layer multimodal Transformer backbone.

[0035] Cross-modal attention sublayer: Using geometric tokens as queries and voxel data tokens as key values, it computes the attention pattern of the geometric token under the current bit conditions through multi-head attention, guiding the geometric features to learn how to encode and reconstruct the bits.

[0036] In this cross-modal attention sublayer, the geometry token is used as the query and the voxel data token as the key. Multi-head attention is then used to compute the geometry token under the current bit conditions.

[0037] In the cross-modal attention mechanism, the geometric token obtained by three-dimensional voxel mesh through three-dimensional convolutional encoding and flattening is represented as the query matrix.

[0038] Q=G∈ The text token obtained by Transformer encoding the structured text consisting of the target binary sequence, voxel resolution, slice axis, and modulus parameters is represented as a key matrix and a numerical matrix.

[0039] K=V=T∈ The query matrix is ​​used to characterize the local geometric features of different spatial locations of the 3D model, and the key matrix is ​​used to characterize the semantic information of the target bits and voxel configuration parameters. Thus, the semantic guidance and fusion of bit and voxel parameters on geometric features are achieved through attention weight calculation.

[0040] QK⊤∈ This represents the similarity between each geometric token and each text token. This is a scaling factor used to suppress numerical instability in the inner product as the dimension increases; dz=D represents the feature dimension of the Query and Key.

[0041] Different attention heads can focus on different segments and different voxel parameter fields in the bit sequence, thereby guiding geometric features to learn how to encode and reconstruct.

[0042] Convolutional Feedforward (CLMP) Sublayer: The updated geometric tokens are rearranged back into the 3D structure and fed into the convolutional multilayer perceptron CLMP. CLMP uses 3×3 and 1×1 convolutions and upsampling structures to aggregate contextual information within a local spatial range; simultaneously, gating parameters are introduced. right Modulate it to aggregate it layer by layer with the model. Related global geometric capabilities and bit matching information.

[0043] After stacking multimodal Transformers and CLMPs, the final CLS corresponding to each point cloud model in the dataset held by the brand can be obtained, denoted as... Ultimately, CLS integrates the geometric information from each model with the conditional information from the text configuration, and uses it as input for the subsequent three types of output heads.

[0044] The three types of output heads include the model selection head: for Perform multi-class classification and output the probability that each model is selected as the adjudication model. This yields the identifier of the prediction model, denoted as model_id. model_id is used only as an index field to locate the corresponding 3D model in the shared model library. During encoding, only the geometric voxel features of the model are represented and learned, and model_id itself is not embedded or its features are extracted.

[0045] Rotation Angle Regression Head: Linear mapping is performed on the CLS representation of the selected model, the neural network regresses 9D rotation, and in post-processing, it is projected to the rotation group SO(3) through Gram-Schmidt orthogonalization.

[0046] Voxel parameter regression: Classifies or regresses the CLS representation, outputting a predicted vector of voxelized parameters. Including predicted voxel resolution Predicted slice axis and the predicted modulus In this process, the combination that best suits the current bit sequence is automatically selected from several alternative resolutions and moduli.

[0047] Loss function and training objective; During the training phase, a joint loss was designed to ensure that the triples output by the network could stably reconstruct the target sequence at the decoding end.

[0048] in, These are the weighting coefficients; Sequence consistency loss This is used to constrain the decoding result to be as consistent as possible with the target sequence; Angle Regression Loss The training samples contain a reference rotation angle. Therefore, a combination of cross-entropy and L2 regularization is used to construct the angle regression loss for the three-axis angles or polar coordinate parameters to improve the accuracy and stability of angle prediction; Angular domain constraint loss Maintain continuity at adjacent boundaries; Voxel parameter selection loss This enables the network to learn to select appropriate voxelization configurations for different bit sequences.

[0049] By minimizing the aforementioned joint loss, the network can automatically output a model ID, three-dimensional rotation angle, and voxelization parameters at the encoder end that can stably reconstruct the target bit sequence at the decoder end.

[0050] To achieve the mapping of "point cloud to binary key", this invention defines the following geometric transformation process, including the following steps: Step 1, Point Cloud Normalization: The input 3D point cloud data is normalized to ensure that all models in the dataset held by the brand are at a uniform degree. Let... Represents the three-dimensional coordinates of the i-th original point, after normalization. The calculation is as follows:

[0051] Among them, the centroid Represents the 3D coordinate vector of the i-th original point cloud point. N is the total number of points in the original point cloud. The centroid of the point cloud, calculated as the average of all point coordinates, is used to center the model. Scaling factor. Used to normalize point cloud coordinates to a unit space This is to prevent numerical overflow or the impact of differences in scale between different models.

[0052] Step 2, rotation processing; The point cloud model is rotated around its centroid based on the input rotation angles. This rotation can be achieved using Euler angles or a 9D representation of the rotation matrix. For a given rotation angle around axes X, Y, and Y, the rotation can be represented by the rotation matrix R. Step 3, orthogonal projection of the 9D rotation matrix; If the rotation angle is represented by 9D, to ensure the effectiveness of the rotation, the Gram-Schmidt orthogonalization method needs to be used to project R onto the SO(3) group. The expression for the rotation matrix is ​​as follows: (3) Where represents the orthogonalization function.

[0053] Then, rotate the point cloud, as expressed by: (4) in, Let represent the 3D coordinates of the i-th element after normalization of the point cloud model.

[0054] Step 4, voxelization; like Figure 2 As shown, the rotated point cloud data is converted into a voxel grid representation for processing by the neural network. A fixed-size voxel grid is selected, and the grid is filled according to the density of the point cloud data. The judgment rule is: if the voxel occupies a density greater than a threshold, the voxel is marked as 1, indicating that the voxel is occupied; otherwise, the voxel is marked as 0.

[0055]

[0056] in, Position in the voxel grid The occupancy state, where V is the volume of the model. If the occupancy of the voxel mesh reaches... ,So =1; The voxel index is calculated based on the rotated x-axis, y-axis, and z-axis coordinates. Indicates voxel labeling.

[0057] Step 5: Slice counting and binary sequence generation; Count the number of voxels occupied in each xy plane along the selected axis direction. :

[0058] Perform a modulo operation on the number of voxels in each slice and generate a binary sequence:

[0059] in, This represents the total number of voxel grids in the w-th layer. After taking the modulus m, convert it to binary. Finally, the binary results of all the slices are concatenated to form the target binary sequence. .

[0060] Step 6: Obtain the binary secret information to be transmitted and the preset voxelization configuration parameters; the voxelization configuration parameters include voxel resolution, slice axis and modulus number; Specifically, the binary sequence transmitted in this invention is an authorization token encapsulated with information. Information such as printer process parameters, printer ID, authorized machine number, delivery date, print count limit, and message authentication code are manually converted into binary encapsulation to generate the target binary sequence to be transmitted.

[0061] Step 7: Combine the binary secret information with the voxelization configuration parameters and generate the index text; Text dataset generation: Traverse the 3D model library, rotating each model along each axis with a 5-degree rotation step at voxel resolution. With a voxelization threshold of 60%, generate a dataset containing indexed text. This index includes the model ID, resolution, and rotation matrix components. Threshold , slice axis and module.

[0062] Step 8: Input the indexed text and candidate 3D models from the shared 3D model library into the pre-trained encoding neural network; Step 8: Using an encoding neural network, based on a cross-modal attention mechanism, the text features of the index text and the geometric features of the candidate 3D model are fused to output the model identifier, 3D rotation angle and optimized voxelization parameters of the target 3D model. The neural network outputs a corresponding triple: given the target bit sequence and candidate voxel rules, the encoding-side neural network outputs a triple (model_id, ...). This is the final encoded result, used by the decoding end. For three-dimensional rotation angle, These are the voxelization parameters.

[0063] Step 9: Send only the model identifier, three-dimensional rotation angle, and voxelization parameters through the public channel so that the receiving end can recover the binary secret information based on the parameters.

[0064] Step 10: The sender transmits only the model identifier, rotation angle, and voxelization parameters in the public channel, without sending any explicitly encrypted 3D model or independent encrypted file. The receiver, possessing the same 3D model library and pre-shared key, loads the target 3D model from its local library based on the model identifier, performs a geometric decoding process according to the agreed geometric transformation and voxelization rules to obtain a binary sequence, and uses the pre-shared key to decrypt the sequence to recover the original secret information.

[0065] like Figure 3 As shown, this embodiment illustrates a complete business application process, including encryption and decryption processes, as detailed below: Encryption process (brand): The brand generates a binary sequence to be transmitted (e.g., compressed encoding of printing process parameters and printer ID). This sequence is then input into the trained neural network to obtain the corresponding geometric parameter triplet.

[0066] Decryption process (3D printing factory): The 3D printing factory receives the encrypted message (geometric parameter triplet) and performs the decryption operation:

[0067] With access to the shared model library and a pre-shared key, the receiving end loads the model based on `model_id`, performs geometric transformations and slice statistics according to steps 1-5 based on the received rotation angle, and recovers the binary sequence. Finally, the recovered sequence is... The accuracy is calculated by comparing the sequence with the actual binary sequence, thus verifying the accuracy of the protection provided by both parties' protocols. If they match, the authorization verification and anti-counterfeiting verification are completed.

Claims

1. A carrier-free information encryption transmission method based on a three-dimensional model, characterized in that, Includes the following steps: S1, the sender generates binary ciphertext based on the binary secret information to be transmitted; S2, the sender combines the binary ciphertext with the corresponding voxelization configuration parameters to generate index text; S3, the text index and the geometric model in the preset 3D model library are input into the pre-trained neural network encoder, which outputs the model identifier, 3D rotation angle, and voxelization parameter triplet of the target 3D model; S4, the receiver, having the same 3D model library and pre-shared key, loads the target 3D model from the local library according to the model identifier, performs the geometric decoding process according to the agreed geometric transformation and voxelization rules, obtains the binary sequence, and decrypts the binary sequence using the pre-shared key.

Images