An anti-interference perception encryption method for unmanned aerial vehicle images

Abstract

This invention discloses an anti-interference perceptual encryption method for UAV images, belonging to the field of image encryption technology. The method constructs a joint deep learning model, including a visual encryption module, a visual decryption module, and a differential attack module based on a bidirectional reversible network. During training, the UAV image is processed through discrete wavelet transform, the bidirectional reversible network, and inverse discrete wavelet transform to generate an encrypted image and key information. The differential attack module simulates interference to generate an interfered encrypted image; this image, along with the key, is then input into the decryption module to recover the original image. The encryption and decryption modules are simultaneously optimized using a composite loss function. This method achieves perceptual encryption and high-fidelity reversible recovery, possesses excellent anti-interference capabilities, is suitable for UAV transmission scenarios, supports real-time processing and high-security decryption, and can be deployed on airborne edge computing platforms or cloud systems, meeting privacy protection requirements.

Description

Technical Field

[0001] This invention relates to the field of image encryption technology, and more specifically to an anti-interference sensing encryption method for UAV images. Background Technology

[0002] Object detection, a key technology in computer vision, is a core component of drone aerial photography data processing. However, while capturing valuable scene information, it inevitably records a large amount of personal privacy data, such as pedestrian facial features, vehicle license plate numbers, and environmental details of sensitive areas. Protecting personal privacy during the use, transmission, and distribution of drone aerial photography data, while simultaneously ensuring data availability and transmission resistance to interference, has become a significant technical challenge.

[0003] Currently, privacy protection methods primarily rely on destructive information anonymization techniques, the core of which is the permanent destruction of original pixel information. This destructive operation results in the complete and irrecoverable loss of clear, verifiable original image information. This irreversibility directly renders such methods ineffective in scenarios requiring legal tracing, judicial evidence collection, or secondary data mining. As an alternative to destructive methods, academia has also explored protection schemes based on traditional cryptography. However, when these methods are ported to the large-scale, dynamic, real-time image processing scenarios of drone aerial photography, the primary challenge is the excessively long processing time. Because traditional encryption algorithms were designed primarily for the confidentiality of text or binary data and were not optimized for the high throughput of aerial visual data from drones, their high-intensity operations at the pixel level consume significant computing resources, making it difficult to meet the stringent requirements of real-time encryption and rapid transmission on the drone's onboard end, easily causing data transmission delays and stuttering. More importantly, drone aerial image transmission is susceptible to interference from factors such as flight altitude, environmental obstruction, and signal frequency bands, which can easily lead to problems such as Gaussian noise, salt-and-pepper noise, and data compression distortion. Traditional destructive privacy protection methods and traditional cryptographic encryption schemes have weak anti-interference capabilities. When encrypted images encounter such interference, the visual information becomes more blurred, or even completely loses its preview and downstream analysis value. The decryption quality will be severely reduced, or even unable to be decrypted normally, making it unsuitable for the actual needs of dynamic drone transmission scenarios. Summary of the Invention

[0004] Based on the above-mentioned technical problems, this invention proposes an anti-interference sensing encryption method for UAV images.

[0005] The technical solution adopted in this invention is: An anti-interference sensing encryption method for UAV images includes the following steps: S1. Construct a joint deep learning model; The joint deep learning model includes a visual encryption module built on a bidirectional reversible network, a visual decryption module for image restoration, and a differentiated attack module to improve anti-interference capabilities. S2. Train the constructed joint deep learning model, as follows: S21. Acquire aerial images taken by the drone and divide them into training and validation sets; S22. Input the images in the training set into the visual encryption module, and process them sequentially through discrete wavelet transform, bidirectional invertible network and inverse discrete wavelet transform to generate encrypted images with perceptual structure and key information for subsequent authorization and decryption. S23. Input the encrypted image into the differential attack module to perturb the encrypted image and generate a perturbed encrypted image; S24. Input the interfered encrypted image and the corresponding key information into the visual decryption module. By performing the reversible transformation of the encryption process in reverse, the interfered encrypted image is decrypted and the decrypted image is output. S25. Determine the composite loss function, and simultaneously optimize the visual encryption module and the visual decryption module during training to obtain a well-trained joint deep learning model. S3. The user inputs the drone aerial images to be processed into the trained joint deep learning model for image encryption and high-fidelity restoration.

[0006] The beneficial technical effects of the present invention are as follows: (1) This invention achieves perceptual encryption and high-fidelity reversible recovery, taking into account both privacy and usability: Unlike traditional destructive privacy protection technologies that permanently destroy pixel information, this invention constructs a bidirectional encryption-decryption mapping based on a bidirectional reversible network to achieve perceptual encryption, meeting the actual needs of rapid preview and retrieval of UAV images; after authorized decryption, the original image can be restored with high fidelity, and is highly consistent with the original image at the pixel and semantic levels, meeting the needs of scenarios such as secondary data mining and judicial evidence collection.

[0007] (2) It has excellent anti-interference capabilities and is suitable for UAV transmission scenarios: By using the differentiated attack module to simulate various interferences in the UAV transmission process during the training phase, the decryption module can still output high-fidelity restored images when the encrypted image encounters interference such as Gaussian noise, compression, and cropping. This solves the problem of weak anti-interference capabilities of traditional solutions and is suitable for complex scenarios of dynamic transmission of UAVs.

[0008] (3) Integrated design to ensure real-time processing: The end-to-end integrated architecture is adopted to deeply integrate the visual encryption module and the visual decryption module. No additional redundant modules are required, which realizes efficient collaboration between encryption and decryption functions and avoids extra time consumption for data interaction between modules. At the same time, it is compatible with the UAV onboard edge computing platform, with low inference latency and high computing efficiency, which can meet the real-time processing needs of UAV aerial images.

[0009] (4) Introducing Discrete Wavelet Transform and Inverse Discrete Wavelet Transform to improve encryption accuracy and anti-interference foundation: This invention integrates Discrete Wavelet Transform (DWT) and Inverse Discrete Wavelet Transform (IWT) in the visual encryption and decryption module, which enhances the spatial-frequency feature modeling capability and works in synergy with the differentiated attack module to further strengthen the model's anti-interference capability.

[0010] (5) With the total loss as the target, the trainable parameters in the visual encryption module and the visual decryption module are optimized through backpropagation, so that the parameters of the visual encryption module and the decryption module are updated synchronously and adjusted in coordination. This ensures that the encrypted image generated by the encryption module can resist interference and can be accurately restored by the decryption module, thus realizing the coordinated adaptation of the encryption-decryption link and taking into account privacy protection, anti-interference and decryption fidelity.

[0011] (6) High security and convenient deployment: With the aid of key information for decryption, the original image cannot be recovered without the correct key. Combined with multi-layer serial reversible network stacking structure, multi-branch RDB residual dense block nonlinear mapping, anti-interference composite loss constraint training and other multi-level encryption design, the key space complexity is extremely high, which greatly increases the difficulty of reverse cracking. The end-to-end integrated architecture does not require manual intervention and can be seamlessly deployed on the airborne system of UAV equipped with edge computing module. It can also be adapted to cloud storage and transmission scenarios, flexibly meeting the privacy protection needs of various UAV aerial photography. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the anti-interference sensing encryption method for UAV images according to the present invention. Figure 2 This is a schematic diagram of the processing flow of the reversible coupled block in the bidirectional reversible network during the encryption phase of this invention. Figure 3 This is a flowchart of the differential attack module in this invention; Figure 4 This is a flowchart of the image processing during encryption / decryption training for this invention. Figure 5 These are visual effect comparison diagrams for specific applications of the present invention. Detailed Implementation

[0013] This invention proposes an image anti-interference perception encryption method for use in UAV aerial photography scenarios. It aims to achieve a balance between UAV image privacy protection, perception availability, and transmission anti-interference by constructing a bidirectional reversible network architecture and integrating an anti-interference training mechanism, thus meeting the practical application needs of UAV inspection, security, and surveying scenarios. Specifically, this invention can encrypt UAV aerial images at the thumbnail level to ensure image privacy and security while meeting the practical needs of rapid image preview and retrieval. It can also decrypt and recover the original image information with high fidelity under authorized conditions. Furthermore, it possesses excellent anti-interference capabilities to adapt to interference scenarios such as noise, compression, and cropping during UAV transmission, thereby achieving a good balance between the strength of privacy protection, perception availability, and information recovery fidelity, and adapting to the actual needs of dynamic UAV transmission.

[0014] like Figure 1 As shown, an anti-interference sensing encryption method for UAV images has the following specific steps: S1. Construct an end-to-end joint deep learning model, which consists of three core modules that work in cascade: a learnable visual encryption module based on a bidirectional reversible network, a visual decryption module for high-fidelity image restoration, and a differentiated attack module to improve anti-interference capabilities. The visual encryption module and the visual decryption module share the same network architecture and parameters, only with opposite data flow directions to achieve bidirectional reversible mapping. The differentiated attack module is used to simulate transmission interference during the training phase.

[0015] S2. During the model training phase, the drone aerial image dataset used for the perception encryption algorithm is randomly divided into a training set and a validation set according to a preset ratio. The resolution of all images is uniformly adjusted to a preset size before being used for model training.

[0016] S3. Input the original image into the visual encryption module, which integrates a bidirectional reversible network, Discrete Wavelet Transform (DWT) and Inverse Discrete Wavelet Transform (IWT) to generate an encrypted image with a perceptual structure, and at the same time generate key information for subsequent authorization and decryption.

[0017] S4. Input the encrypted image generated in S3 into the differential attack module. This module randomly simulates common interference scenarios during drone image transmission (including Gaussian noise, salt and pepper noise, JPEG compression, and random cropping) to perturb the encrypted image and generate an interfered encrypted image.

[0018] S5. Input the interfered encrypted image and the corresponding key information into the visual decryption module. This module decrypts the interfered encrypted image by performing the reversible transformation of the encryption process in reverse and outputs a high-fidelity restored image that is highly similar to the original image.

[0019] S6. Define a composite loss function to jointly optimize the parameters of each module of the model, achieving a synergistic improvement in encryption effect, decryption fidelity, and anti-interference capability. This composite loss function includes encryption loss, decryption loss, and structural similarity loss. The encryption loss constrains the consistency between the encrypted image and the target encrypted image to ensure perceptual usability. The target encrypted image is generated by uniformly dividing the original image into blocks of corresponding block sizes and replacing all pixel values ​​in each block with the average pixel value within that block, serving as the training reference target for the encryption process. The decryption loss constrains the consistency between the restored image and the original image to ensure high-fidelity restoration. The structural similarity loss constrains image consistency from three dimensions: brightness, contrast, and local structure, helping to maintain the spatial coherence and semantic integrity of the image.

[0020] S7. During training, based on the composite loss function calculated in S6, the learnable parameters of the visual encryption module and the visual decryption module are simultaneously optimized through the backpropagation algorithm, so that the model can achieve bidirectional reversibility of encryption and decryption while ensuring the strength of privacy protection, perceptual usability and anti-interference ability.

[0021] S8. In the application phase, the user inputs the drone aerial images to be processed into the trained joint deep learning model. The model first inputs the images into the visual encryption module to encrypt the images, generating encrypted images that can be previewed, transmitted, and stored. The key information is saved separately. When it is necessary to restore the original image, the authorized user can send the encrypted image and the corresponding key information into the visual decryption module to achieve high-fidelity restoration of the original image, meeting the needs of subsequent analysis, archiving, or judicial evidence collection.

[0022] The core of this invention is an end-to-end joint encryption and decryption model, namely a joint deep learning model. This model is based on a bidirectional reversible network architecture, deeply integrates visual encryption and decryption modules, and introduces a differentiated attack module. It relies on Discrete Wavelet Transform (DWT) and Inverse Discrete Wavelet Transform (IWT) to achieve high-fidelity encryption and decryption without the need for additional redundant modules, and is adapted to the real-time processing requirements of UAV onboard terminals.

[0023] The following provides a detailed explanation of each module and its processing flow.

[0024] 1. Visual encryption module; This module is responsible for applying recoverable visual transformations to the target image, achieving a balance between privacy masking and perceptual usability. Its input is the original drone aerial image, and the core workflow revolves around DWT, IWT, and bidirectional reversible networks, as detailed below: Discrete Wavelet Transform (DWT) Processing: First, process the original UAV image... A multi-scale frequency domain decomposition (DWT) is performed to decouple the image into a low-frequency (LL) component and three high-frequency components (HL, LH, and HH). The LL component preserves the global contour and basic perceptual structure of the image, ensuring the encrypted image has a clear thumbnail preview capability. The high-frequency components carry sensitive details such as image edge texture, pedestrian faces, and license plates, serving as the core perturbation objects for visual encryption. The LL, HL, LH, and HH frequency band features obtained from the decomposition are concatenated through channel replication to form an 8-channel feature sequence containing two complete sets of frequency domain information. This sequence is then integrated along the channel dimension and input into a subsequent reversible network, thus balancing the preservation of global structure with the encryption requirements of high-frequency sensitive information.

[0025] Reversible network transformation: The spliced ​​full-frequency domain features are sequentially processed through 12 layers of serially stacked reversible coupled blocks. Features are transferred between blocks in a serial cascade manner, without parallel computation. Each reversible coupled block uses a residual dense block (RDB) as its core computational unit. The overall processing flow is as follows: Figure 2 As shown: The input is equally divided into two feature branches, X1 and X2, both containing complete LL low-frequency and HL, LH, and HH high-frequency wavelet information. X1 is connected to the first residual dense block, i.e., the first RDB, focusing on extracting the low-frequency global contour structure to maintain the basic visual integrity of the image. X2 is added and fused with the features output from the uplink branch (i.e., the first RDB) and then fed into the second residual dense block (i.e., the second RDB), the third residual dense block (i.e., the third RDB), and the stitching module. In the second RDB, the nonlinear perturbation and encryption transformation of high-frequency texture details are focused. The branch features are multiplied and affinely coupled with X1 after exponential nonlinear operation, and then added and fused with the long-distance residual compensation features output from the bottom third RDB. Finally, the single-block reversible transformation is completed by stitching through the Concat channel (i.e., the stitching module). In the reversible coupling block, the RDB efficiently captures multi-scale frequency domain contextual correlations through multi-layer dense residual connections, while satisfying the constraints of the reversible coupling structure, providing a foundation for subsequent high-fidelity decryption. During the encryption phase, the frequency domain features pass through the reversible coupling block layer by layer in a forward order from 1 to 12. The correlation between high and low frequency features is disrupted through multi-path coupling operations, thus completing the visual concealment of the image.

[0026] Inverse Discrete Wavelet Transform (IWT) and Key Generation: After the reversible network completes 12 layers of serial coupling transformation, the features of the X1 and X2 branches are mapped to the spatial domain via Inverse Discrete Wavelet Transform (IWT). The X1 branch generates the encrypted image for public transmission. The X2 branch is reconstructed into a key feature image, i.e., key information, which serves as the core basis for decryption authorization. This key is used throughout the entire encryption and decryption process. Without the correct key, it is impossible to reverse the coupling operation of the reversible network and the inverse wavelet transform, and it is impossible to restore the original frequency domain features and spatial domain image, thus effectively ensuring the privacy and security of UAV images.

[0027] Encryption Constraints: An encryption loss is introduced during the encryption process to constrain the encrypted image. Visual consistency with the target encrypted image ensures both high visual quality and perceptual usability of the encrypted image.

[0028] 2. Differentiated attack module; like Figure 3 As shown, to improve the model's anti-interference capability during UAV image transmission, this differentiated attack module is introduced between the visual encryption module and the visual decryption module during the training phase. This simulates common interference scenarios in UAV image transmission, allowing the model to optimize parameters under perturbed conditions. The specific interference types and settings are as follows: Gaussian noise: To simulate sensor noise or transmission channel interference, Gaussian noise with zero mean and random variance (0≤std≤0.5) is added to simulate detail distortion in actual transmission. Salt and pepper noise: simulates pixel distortion during transmission, replacing pixels with extreme values ​​(0 or 255) with a probability of 0.01 to 0.5, and maintaining differentiability through smooth approximation; JPEG compression: Image compression in analog transmission, using a quality factor of 20~90 for compression-reconstruction, and employing a continuous quantization approximation strategy to achieve differentiability for non-differentiable operations such as quantization and rounding.

[0029] Random cropping: Simulating partial information loss during transmission, the encrypted image is locally cropped and resampled at a ratio of 0.1 to 0.9. This is achieved through continuous cropping based on interpolation, which enables gradients to be effectively propagated.

[0030] This module embeds the aforementioned interference into the training computation graph and uses a composite loss function to iteratively optimize the model parameters. This allows the decryption module to stably output high-fidelity restored images even in the presence of interference, significantly improving the model's robustness against interference. The specific parameter optimization steps are as follows: (1) Perturbation application: The visual encryption module outputs an encrypted image. Then, the differential attack module randomly applies one of the aforementioned random strengths of interference to generate a jammed encrypted image. The interference only affects the encrypted image and does not affect the key information.

[0031] (2) Decryption reasoning: decrypt the interfered encrypted image The key image, or key information, is input into the visual decryption module, which outputs the decrypted and restored image. .

[0032] (3) Loss calculation: On the one hand, calculate the decrypted and restored image. With original drone images To address pixel-level errors and visual perceptual similarity, and on the other hand, to introduce encryption-side constraint loss, the decryption loss and encryption loss are weighted and fused to construct the total loss. The model reconstruction bias and fidelity error are jointly quantified from both the encryption and decryption ends.

[0033] The formula for calculating the total loss is as follows: ; In the formula: These are the manually preset weight hyperparameters for each loss term, which remain fixed during training; For encryption loss, To decipher the loss, These are the encryption structure similarity loss and the decryption structure similarity loss, respectively.

[0034] (4) Gradient backpropagation: The total loss Backpropagation is performed to all trainable weights in the reversible coupled network layers and RDB residual dense blocks within the visual encryption module and visual decryption module. The differential attack module does not participate in parameter updates and only provides perturbation constraints.

[0035] (5) Parameter update: An adaptive optimizer (Adam) is used to iteratively correct the trainable parameters of the encryption and decryption modules based on the gradient information obtained from backpropagation. In other words, the trainable weights and biases inside the neural networks of the visual encryption and decryption modules are iteratively optimized to minimize the total loss. .

[0036] (6) Iterative convergence: Repeat the above process, and through multiple rounds of training, allow the model to adaptively learn interference features and countermeasures, gradually optimize parameters until the model converges, and achieve high-fidelity decryption in interference scenarios.

[0037] 3. Visual decryption module; This module is the core of achieving high-fidelity restoration of the original image. It shares the same network architecture and parameters as the visual encryption module, only the data flow and operation logic are reversed, achieving strict bidirectional reversibility (in the decryption stage, frequency domain features pass through the reversible coupling block layer by layer in reverse order from 12 to 1, and the processing flow in the reversible coupling block is the same as in the encryption stage). The core relies on DWT, IWT, and the bidirectional reversible network. The specific process is as follows: Frequency domain conversion: Converting encrypted images that may be subject to interference after transmission. The corresponding key information is then subjected to DWT transformation to convert it into frequency domain features. These frequency domain features contain interference signals, which need to be suppressed and restored in subsequent steps.

[0038] Reversible transformation restoration: The frequency domain features are inverted and input into a reversible network block, such as... Figure 4As shown, the nonlinear transformations in the encryption process are gradually reversed. In the decryption phase, the process strictly follows the reverse order from 12 to 1, passing through the reversible coupling blocks layer by layer, and performing each mathematical operation in reverse to restore the exponential transformation, multiplicative coupling, residual addition, and dense feature mapping, thus completing the visual reconstruction of the image.

[0039] from Figure 4 As can be seen, in the visual encryption module, the data is processed sequentially through the (i-1)th reversible coupling block, the ith reversible coupling block, and the (i+1)th reversible coupling block. In the visual decryption module, the data is processed in reverse order through the (i+1)th reversible coupling block, the ith reversible coupling block, and the (i-1)th reversible coupling block.

[0040] Spatial domain mapping: All frequency domain features after reconstruction are subjected to IWT transform and mapped back to the spatial domain to generate a high-fidelity restored image, i.e., the decrypted and restored image. The image is the same as the original image. It is highly consistent at the pixel and semantic levels, which can meet the needs of subsequent analysis, archiving, and judicial evidence collection.

[0041] Decryption Constraints: Introduce a decryption loss to constrain the decrypted and reconstructed image. With the original image The pixel and semantic consistency are combined with structural similarity loss to ensure the structural integrity of the restored image and avoid image structural distortion caused by wavelet transform or interference.

[0042] The core of the training process of the joint deep optimization model in this invention revolves around encryption-interference-decryption, achieving end-to-end joint optimization of the parameters of each module. The specific steps are as follows: 1. Data preparation and preprocessing: Select a drone aerial image dataset, randomly divide it into training set and validation set, and uniformly adjust the resolution to ensure that the image size input to the model is consistent.

[0043] 2. Encryption / decryption training steps: For a batch of input original images First, the visual encryption module is input, and an encrypted image is generated through DWT transformation, reversible network transformation, and IWT transformation. and key information; then encrypt the image. Input the differential attack module, randomly apply a perturbation of random strength, and generate a perturbed encrypted image. Finally, the encrypted image that has been interfered with... The key information, along with the input key, is used to generate the visual decryption module, which then generates the restored image, i.e., the decrypted and restored image. The overall image processing workflow is as follows: Figure 4 As shown.

[0044] 3. Composite Loss Function Calculation: A multi-task joint optimization strategy is adopted to construct a composite loss function, achieving a synergistic improvement in encryption effectiveness, decryption fidelity, and anti-interference capability. The total loss function is as follows: ; in, These are the weighted hyperparameters for each loss term (e.g., set to 4.0, 1.0, 0.5, 0.5 respectively). The specific values ​​can be adjusted according to the actual needs of the drone aerial photography scenario to ensure that parameter optimization balances privacy protection, perceptual usability, and anti-interference capabilities. The definitions of each loss term are as follows: (1) Encryption loss ( The algorithm consists of pixel-level L1 loss and perceptual loss. The perceptual loss utilizes a pre-trained VGG network to extract high-level semantic features. By comparing feature differences in the perceptual space, it further optimizes the structural and texture similarity of the image. (Constrained encrypted image) Encrypted image of the target This ensures the consistency and usability of encrypted images while preventing the leakage of sensitive information. The specific formula is as follows: ; in, This represents high-level semantic features extracted by a pre-trained VGG network. For pixel-level L1 loss, These are the weighting coefficients (e.g., set to 1.0, 0.2).

[0045] (2) Decryption loss ( Using the same pixel-perceptual joint loss as the encryption stage, the constrained image recovery is equivalent to the decrypted image recovery. With original drone images This design ensures pixel-level and semantic consistency, encouraging the network to maintain consistency in both the spatial and semantic domains, thereby improving recovery accuracy under attack-resistant conditions. The specific formula is as follows: ; in, These are the weighting coefficients, which are exactly the same as the weighting coefficients in the encryption loss (e.g., set to 1.0, 0.2).

[0046] (3) Structural similarity loss Relying solely on pixel errors may lead to insufficient structural information recovery. Therefore, structural similarity loss is introduced in both the encryption and decryption stages to constrain the image structural consistency during encryption and decryption, avoiding image structural distortion caused by DWT transform, IWT transform, and interference, and ensuring the continuity of image context. The specific formula is as follows: ; ; Since the Structural Similarity Index (SSIM) ranges from [0,1], a larger value indicates a higher degree of structural similarity. Therefore, 1−SSIM is used as the minimization objective. The encrypted image has been compromised.

[0047] (4) Backpropagation and parameter update: based on the calculated total loss The gradients of all trainable parameters in the visual encryption and decryption modules are calculated synchronously using the backpropagation algorithm, and the model parameters are updated synchronously. The specific parameter update steps are as follows: (41) Gradient calculation: With the total loss L as the target, the backpropagation algorithm is used to trace back from the output of the decryption module and calculate the gradient of all trainable parameters (reversible coupled block, RDB related parameters) in the decryption module and encryption module in turn to quantify the influence of each parameter on the total loss.

[0048] (42) Gradient correction: Gradient clipping strategy is adopted to suppress gradient explosion, ensure that the gradient value is within a reasonable range, ensure the stability of parameter updates, and avoid model training oscillation.

[0049] (43) Parameter iteration: The Adam adaptive optimizer is used. Combined with the calculated gradient information, the update step size of each trainable parameter is adaptively adjusted. The weights of the reversible coupled block, the RDB convolution kernel and the bias are iteratively updated according to the preset learning rate to gradually reduce the total loss L.

[0050] (44) Collaborative optimization: The parameters of the encryption module and the decryption module are updated synchronously and adjusted collaboratively to ensure that the encrypted image generated by the encryption module can resist interference and can be accurately restored by the decryption module, realizing the collaborative adaptation of the encryption-decryption link, taking into account privacy protection, anti-interference and decryption fidelity.

[0051] (5) Model evaluation and selection: After each training cycle, the performance of the model is evaluated on an independent validation set. Peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are used as objective image quality evaluation criteria. Finally, the model weight with the best overall performance on the validation set is selected as the final deployed model to ensure the stability and reliability of the model in practical applications.

[0052] In other words, the training process iteratively updates the model weights with the goal of minimizing the total loss. However, the weights with the smallest loss in the training set are prone to overfitting and are not the best in terms of generalization ability. Therefore, the weights with the best overall performance can be selected as the final deployment model parameters by screening the image quality index on the validation set.

[0053] After the joint deep optimization model is trained, it can be seamlessly deployed on the UAV-borne edge computing platform. The actual application process is divided into two stages: encryption (UAV-borne end) and decryption (authorization end), as detailed below: Step 1: Encryption (on the drone's onboard device); The drone acquires aerial images to be processed and inputs them into the deployed joint model. The visual encryption module encrypts the images through DWT transformation, reversible network transformation, and IWT transformation to generate encrypted images. and key information; encrypted images For public release, transmission, or storage, key information is stored separately using hardware encryption. Plaintext storage or arbitrary transmission is strictly prohibited to prevent key leakage.

[0054] Step 2: Decrypt as needed (authorized end); In authorized scenarios, encrypted images that may be subject to interference after transmission will be protected. The corresponding key information is input into the model; the visual decryption module performs reversible decryption of the interfered encrypted image through DWT transformation, reversible network inverse transformation, and IWT transformation, restoring the frequency domain features of the original image, suppressing interference signals, and generating a high-fidelity restored image, i.e., the decrypted and restored image. .

[0055] This invention achieves privacy and security protection, perceptually controllable preview, and high-fidelity decryption and recovery of sensitive images captured by drones, while also possessing excellent anti-interference capabilities and adapting to complex scenarios during drone image transmission. The method deeply integrates visual encryption and visual decryption modules into a bidirectional reversible network, introduces a differentiated attack module to simulate transmission interference, and performs end-to-end joint training through a composite loss function, achieving a balance between privacy protection, perceptual usability, and anti-interference capabilities.

[0056] The high-fidelity restored image is visually almost indistinguishable from the original aerial image, such as... Figure 5 As shown in the comparative experiment results, all high-frequency texture information, such as building outlines, road markings, vegetation boundaries, and vehicle details, of the original aerial image can be accurately restored. Even after being subjected to various real-world transmission interferences, including Gaussian noise, salt-and-pepper noise, lossy JPEG compression, and random region cropping, the decrypted image still maintains extremely high visual clarity, with no obvious blurring, distortion, color shift, or structural damage. In terms of quantitative indicators, the PSNR of the decrypted image is higher than 53dB under various interference scenarios, and the SSIM remains stable at 0.9986, with pixel accuracy and structural similarity approaching the level of the original image. This high-fidelity restoration effect can meet the needs of various authorized business applications such as intelligent analysis of aerial images, long-term archiving, and judicial evidence collection. While strictly protecting sensitive privacy information in the image, it fully preserves the usability of the aerial data, achieving a two-way unity between privacy security and business applications.

[0057] To verify the airborne performance of the method of this invention on a practical UAV platform, the Allspark2 scientific research UAV equipped with an NVIDIA Jetson Orin NX high-performance edge computing module was selected as the test platform. This computing platform can provide 100 TOPS (INT8, Sparse) of peak AI computing power, suitable for UAV-based real-time inference scenarios. A desktop platform equipped with an Intel Core i7-14700KF processor and an NVIDIA GeForce RTX 4090D discrete graphics card was used as a benchmark. When running the model of this invention directly on this benchmark platform, the average single-frame encryption inference time was measured to be 133ms. During the airborne deployment phase, after converting the model of this invention to the TensorRT engine and optimizing it with INT8 quantization, the average single-frame encryption inference time of the model on the Jetson Orin NX platform was reduced to 19ms (approximately 53 FPS), which meets the performance requirements of UAV-based real-time visual encryption tasks. Experimental results show that the method of this invention can efficiently adapt to high-performance edge computing platforms, achieving low-latency, high-real-time image encryption inference on the airborne end, fully meeting the performance requirements of actual UAV deployment.

[0058] In summary, this invention provides a complete, end-to-end technical solution for recoverable privacy protection of UAV images based on discrete wavelet transform and bidirectional reversible networks, while also considering both perception availability and transmission anti-interference capabilities. This method effectively solves the problems of slow processing speed and poor anti-interference in traditional encryption technologies, achieving significant improvements in encryption security, decryption fidelity, perception availability, operational efficiency, and anti-interference robustness. Actual deployment and experimental results demonstrate that this solution can stably meet the practical needs of UAVs in scenarios such as inspection, security, and surveying for real-time encryption, efficient transmission, and authorized high-fidelity recovery of aerial images, possessing high practical value and promising application prospects.

Claims

1. An anti-interference sensing encryption method for UAV images, characterized in that... Includes the following steps: S1. Construct a joint deep learning model; The joint deep learning model includes a visual encryption module built on a bidirectional reversible network, a visual decryption module for image restoration, and a differentiated attack module to improve anti-interference capabilities. S2. Train the constructed joint deep learning model, as follows: S21. Acquire aerial images taken by the drone and divide them into training and validation sets; S22. Input the images in the training set into the visual encryption module, and process them sequentially through discrete wavelet transform, bidirectional invertible network and inverse discrete wavelet transform to generate encrypted images with perceptual structure and key information for subsequent authorization and decryption. S23. Input the encrypted image into the differential attack module to perturb the encrypted image and generate a perturbed encrypted image; S24. Input the interfered encrypted image and the corresponding key information into the visual decryption module. By performing the reversible transformation of the encryption process in reverse, the interfered encrypted image is decrypted and the decrypted image is output. S25. Determine the composite loss function, and simultaneously optimize the visual encryption module and the visual decryption module during training to obtain a trained joint deep learning model. S3. The user inputs the drone aerial images to be processed into the trained joint deep learning model for image encryption and high-fidelity restoration.

2. The anti-interference sensing encryption method for UAV images according to claim 1, characterized in that, In S1: Both the visual encryption module and the visual decryption module include multi-layered stacked reversible coupling blocks, and the data flow in the visual encryption module and the visual decryption module is opposite; the reversible coupling block includes several residual dense blocks.

3. The anti-interference sensing encryption method for UAV images according to claim 2, characterized in that, S22 includes the following steps: S221. Perform discrete wavelet transform processing on the image: First, decompose the image into low-frequency components and high-frequency components by multi-scale frequency domain decomposition; the frequency band features of the low-frequency components and high-frequency components are copied and spliced ​​through channels to form a channel feature sequence containing complete frequency domain information, and then integrated along the channel dimension to form a full-frequency domain feature input reversible network. S222. Perform reversible network transformation: The full-frequency domain features are processed step by step through multiple stacked reversible coupling blocks, and the features are transferred between the reversible coupling blocks in a serial cascade manner. The processing flow of a single reversible coupling block is as follows: The input is equally divided into two feature branches, X1 and X2, both of which contain complete low-frequency and high-frequency wavelet information; X1 is connected to the first residual dense block, i.e., the first RDB. The features processed by the first RDB are added and fused with X2. The fused features are then sent to the second residual dense block, i.e., the second RDB, the third residual dense block, i.e., the third RDB, and the splicing module, respectively; the branch features processed by the second RDB are then subjected to exponential nonlinear operation and multiplicative affine coupling with X1, and then added and fused with the long-distance residual compensation features output by the third RDB. Finally, the single-block reversible transformation is completed by splicing through the splicing module, and the spliced ​​features are output. S223, Inverse Discrete Wavelet Transform Processing and Key Generation: The features output by the splicing module are mapped to the spatial domain via inverse discrete wavelet transform; wherein the X1 branch generates the encrypted image for public transmission. The X2 branch is reconstructed into key information, which serves as the core basis for decryption authorization.

4. The anti-interference sensing encryption method for UAV images according to claim 3, characterized in that, In S23: During the training phase, a differential attack module is introduced between the visual encryption module and the visual decryption module to simulate interference scenarios in image transmission, enabling the joint deep learning model to optimize parameters under perturbation. Specific interference scenario types include: Gaussian noise, salt-and-pepper noise, JPEG compression, and random cropping.

5. The anti-interference sensing encryption method for UAV images according to claim 4, characterized in that, S24 includes the following steps: S241, the encrypted image that has been interfered with. The corresponding key information is subjected to discrete wavelet transform to convert it into frequency domain features. S242. Input the frequency domain features into the multi-layer stacked reversible coupling block in reverse to perform reverse restoration, complete the image visual restoration, and obtain the restored frequency domain features. S243. Perform inverse discrete wavelet transform on the restored frequency domain features to map them back to the spatial domain and generate the decrypted and restored image. .

6. The anti-interference sensing encryption method for UAV images according to claim 5, characterized in that: Introduce encryption loss during the visual encryption module processing to constrain the encrypted image. Visual consistency with the target encrypted image; A decryption loss is introduced during the visual decryption module processing to constrain the decrypted and restored image. With the original image The pixel and semantic consistency are ensured; at the same time, structural similarity loss is introduced in the visual encryption module and visual decryption module respectively to ensure the structural integrity of the decrypted and restored image.

7. The anti-interference sensing encryption method for UAV images according to claim 6, characterized in that: The differential attack module embeds the interference scene into the training and combines iterative optimization of parameters with a composite loss function, enabling the visual decryption module to stably output a high-fidelity restored image even in the presence of interference. The parameter optimization steps are as follows: S251, Visual encryption module outputs encrypted image. Then, the differential attack module randomly applies interference to generate a compromised encrypted image. ; S252, the encrypted image that has been interfered with. Input the key information into the visual decryption module, and output the decrypted and restored image. ; S253. Based on the composite loss function, calculate the decrypted and restored image. With original drone images The pixel-level error and visual perception similarity are calculated, and an encryption-end constraint loss is introduced. A weighted fusion is then used to construct the total loss. ; S254, Total Loss Backpropagation is performed to all trainable weights in the visual encryption and visual decryption modules; S255. An adaptive optimizer is used to iteratively adjust the parameters in the visual encryption module and the visual decryption module based on the gradient information from backpropagation, so as to minimize the total loss. S256. Repeat the above process, and through multiple rounds of training, allow the model to adaptively learn interference features and countermeasures, and gradually optimize the parameters until the model converges.

8. The anti-interference sensing encryption method for UAV images according to claim 7, characterized in that: The formula for calculating the total loss function is as follows: ； In the formula: These are the weight hyperparameters for each loss; The encryption loss is calculated using the following formula: ； In the formula, To encrypt the image, Encrypt the target image; This represents high-level semantic features extracted by a pre-trained VGG network. For pixel-level L1 loss, These are weighting coefficients; The decryption loss is calculated using the following formula: ； In the formula, To decrypt and restore the image, Original image; The encryption structure similarity loss and decryption structure similarity loss are respectively calculated using the following formulas: ; ； In the formula, To encrypt the image, For the encrypted image that has been compromised, To decrypt and reconstruct the image, SSIM stands for Structural Similarity Index.

9. The anti-interference sensing encryption method for UAV images according to claim 8, characterized in that, In S3: After the joint deep learning model is trained, it is deployed on the UAV-borne edge computing platform. The actual application process is divided into two stages: UAV-borne encryption and authorized decryption, as detailed below: S31, Unmanned Aerial Vehicle (UAV) onboard encryption; The drone acquires images to be processed and inputs them into a pre-deployed joint deep learning model. The visual encryption module encrypts the images using discrete wavelet transform, invertible network transform, and inverse discrete wavelet transform, generating encrypted images and key information. The encrypted images are used for public release, transmission, or storage, while the key information is stored separately using hardware encryption. S32. Authorized end decrypts on demand; In authorized scenarios, the transmitted, distorted, encrypted image and the corresponding key information are input together into the joint deep learning model; The visual decryption module uses discrete wavelet transform, inverse reversible network transform, and inverse discrete wavelet transform to reversibly decrypt the interfered encrypted image, restore the frequency domain features of the original image, suppress interference signals, and generate a high-fidelity decrypted and restored image.

Images