Image classification system for personal privacy protection, and method therefor
The image classification system uses photon counting and dual random phase encoding with deep learning to encrypt and classify images, addressing data access vulnerabilities and enhancing security and efficiency.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- DAEGU GYEONGBUK INSTITUTE OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2025-05-27
- Publication Date
- 2026-05-07
AI Technical Summary
Existing data protection methods for cloud-based data analysis are inadequate as they allow access to original data by attackers or cloud service providers, despite encryption, necessitating a solution that encrypts and classifies images to protect personal privacy.
An image classification system using photon counting imaging and dual random phase encoding, combined with deep learning models, generates photon-limited and encrypted images, training classification models through federated learning and adaptive transfer learning to classify images while keeping original data secure.
Enhances security by preventing access to original data, reduces storage needs, and allows efficient use on devices with limited memory, while maintaining high classification accuracy.
Smart Images
Figure KR2025007189_07052026_PF_FP_ABST
Abstract
Description
Video classification system and method for protecting personal privacy
[0001] The present invention relates to an image classification system and method for protecting personal privacy, and more specifically, to an image classification system and method for protecting personal privacy that protects personal information within an original image using Photon counting imaging (PCI) or Double random phase encoding (DRPE) and classifies it through a deep learning model.
[0002] Recently, the importance of data is increasing as vast amounts of data collected using multiple digital devices are being utilized across various industries. For example, cloud computing platforms such as Google, Amazon, and Microsoft provide scalable storage space and computing power to those who wish to analyze data but lack the financial means to purchase infrastructure.
[0003] However, there are limitations to privacy protection and security because data outsourced to cloud servers can be stolen by attackers or accessed by cloud service providers.
[0004] In response to this, although technologies using encryption algorithms to transform data into various forms have been researched to protect data outsourced to cloud servers, cloud service providers and attackers can access the original data because it must be decrypted back to the original before analysis is performed.
[0005] Therefore, there is a need for technology that encrypts original data and classifies encrypted images to protect personal privacy within the data.
[0006] The technology forming the background of the present invention is described in Korean Registered Patent No. 10-2535810 (published May 23, 2023).
[0007] As such, the present invention aims to provide an image classification system and method for protecting personal privacy by using photon counting imaging or dual random phase encoding to protect personal information within a photon source image and classifying it through a deep learning model.
[0008] According to an embodiment of the present invention for achieving such technical challenges, an image classification system for protecting personal privacy comprises: an image processing unit that generates at least one of a photon-restricted image, a photon-restricted cryptographic image, and a cryptographic photon-restricted image using a preset method according to a user's selection on an original image stored in a cloud or on a personal device; and a classification unit that classifies an image using at least one of a first to third classification model learned on at least one of the photon-restricted image, the photon-restricted cryptographic image, and the cryptographic photon-restricted image of the generated personal device, and further comprises a learning unit that trains the first to third classification models to classify an image using a plurality of photon-restricted images, photon-restricted cryptographic images, and cryptographic photon-restricted images, wherein the original image for prior training of the classification model is stored in the cloud, and the original image to be classified by the user may be stored in the personal device.
[0009] The image processing unit above can generate a photon-limited image by applying Poisson multinomial distribution-based photon counting imaging (PMD-PCI) to the original image stored in the cloud or personal device.
[0010] The above learning unit can use initial weights transmitted from a central server and repeatedly train the first classification model for a number of local epochs, update the weights of the first classification model using the learning rate and loss of the local epochs, and transmit the trained weights of the first classification model to the central server to perform weight updates.
[0011] The above classification unit can classify images according to the output class by applying the first classification model learned to the photon-limited image of the generated personal device.
[0012] The image processing unit can generate a photon-limited cryptographic image by sequentially applying Double Random Phase Encoding (DRPE) and Multispectral Photon Counting Imaging (MPCI) to the original image stored in the cloud or personal device.
[0013] The image processing unit above can generate a encrypted photon-limited image by sequentially applying multispectral photon counting imaging and dual random phase encoding to an original image stored in the cloud or personal device.
[0014] The above learning unit performs learning to classify images of a first dataset composed of photon-limited cryptographic images or cryptographic photon-limited images using adaptive transfer learning, and performs re-learning to classify images of a second dataset composed of photon-limited cryptographic images or cryptographic photon-limited images to train the second classification model, wherein the first dataset is a large-scale dataset and the second dataset may be a dataset composed of target data to be classified.
[0015] The above classification unit can classify images according to the output class by applying the second classification model learned to the photon-limited cipher image or the cipher photon-limited image of the generated personal device.
[0016] The above learning unit trains a teacher model to classify images of a second dataset composed of photon-limited cryptographic images or cryptographic photon-limited images using knowledge distillation (KD), and trains a third classification model by training a student model using the second dataset and the output of the teacher model, wherein the second dataset may be a dataset composed of target data to be classified.
[0017] The above classification unit can classify images according to the output class by applying the third classification model learned to the photon-limited cipher image or the cipher photon-limited image of the generated personal device.
[0018] In an image classification method for protecting personal privacy according to another embodiment of the present invention, the method comprises: a step in which an image processing unit generates at least one of a photon-restricted image, a photon-restricted encrypted image, and a encrypted photon-restricted image using a preset method according to a user's selection on an original image stored in a cloud or a personal device; and a step in which a classification unit classifies an image using at least one of a first to third classification model learned on at least one of the photon-restricted image, the photon-restricted encrypted image, and the encrypted photon-restricted image of the generated personal device, and further comprises a step in which a learning unit trains the first to third classification models to classify an image using a plurality of photon-restricted images, photon-restricted encrypted images, and encrypted photon-restricted images, wherein the original image for prior training of the classification model is stored in the cloud, and the original image to be classified by the user may be stored in the personal device.
[0019] As such, according to the present invention, large amounts of data can be encrypted at high speed by using an optical-based image processing and encryption technique capable of parallel processing.
[0020] In addition, the security of protecting personal information can be enhanced by combining photon counting with digit-based random phase encoding.
[0021] In addition, by using a lightweight model, the technology can be utilized on devices with high memory efficiency as well as in the cloud.
[0022] In addition, it can prevent attackers or cloud service providers from accessing the original video and save storage space.
[0023] FIG. 1 is a configuration diagram of an image classification system for protecting personal privacy according to one embodiment of the present invention.
[0024] FIG. 2 is a configuration diagram of an image classification system using a first classification model according to an embodiment of the present invention.
[0025] FIG. 3 is a flowchart of an image classification method for protecting personal privacy according to another embodiment of the present invention.
[0026] FIG. 4 is a schematic diagram illustrating an image classification method using a first classification model according to another embodiment of the present invention.
[0027] FIG. 5 is a schematic diagram illustrating an image classification method using a second classification model according to another embodiment of the present invention.
[0028] FIG. 6 is a schematic diagram illustrating an image classification method using a third classification model according to another embodiment of the present invention.
[0029] FIG. 7 is a diagram illustrating the classification accuracy of a first classification model according to another embodiment of the present invention.
[0030] Figure 8 is a graph of the compression ratio according to the expected number of incident photons.
[0031] Figure 9 is a graph of the size reduction rate according to the expected number of incident photons.
[0032] Figure 10 is a graph of classification accuracy according to the expected number of particles of an image encrypted with double random phase encoding and a photon-limited encrypted image according to another embodiment of the present invention.
[0033] Preferred embodiments according to the present invention will be described in detail below with reference to the attached drawings. In this process, the thickness of lines or the size of components shown in the drawings may be exaggerated for clarity and convenience of explanation.
[0034] Furthermore, the terms described below are defined in consideration of their functions within the present invention, and these may vary depending on the intent or practice of the user or operator. Therefore, the definitions of these terms should be based on the content throughout this specification.
[0035] In the embodiments described below, the image classification system (100) for protecting personal privacy is described by specific examples as being performed by a computing device comprising one or more processors and one or more memories capable of performing the above-mentioned process.
[0036] FIG. 1 is a configuration diagram of an image classification system for protecting personal privacy according to one embodiment of the present invention, and FIG. 2 is a configuration diagram of an image classification system using a first classification model according to one embodiment of the present invention.
[0037] As illustrated in FIGS. 1 and 2, an image classification system (100) for protecting personal privacy may include an image processing unit (110), a learning unit (120), and a classification unit (130).
[0038] First, the image processing unit (110) can generate at least one of a photon-limited image, a photon-limited encrypted image, and an encrypted photon-limited image by using a preset method (e.g., Poisson multinomial distribution-based photon counting imaging (PMD-PCI), Multispectral photon counting imaging (MPCI), Double random phase encoding (DRPE)) on an original image stored in the cloud (1) or a personal device (3) according to the user's selection. Here, the cloud (1) stores original images (e.g., open datasets) for pre-training a classification model, and the personal device (3) stores original images that the user intends to classify.
[0039] According to one embodiment of the present invention, the image processing unit (110) can generate a photon-limited image by applying photon counting imaging based on a Poisson multinomial distribution to an original image stored in a cloud (1) or a personal device (3) according to the user's selection.
[0040] According to one embodiment of the present invention, an image processing unit (110) can generate a photon-limited encrypted image by sequentially applying dual random phase encoding and multispectral photon counting imaging to an original image stored in a cloud (1) or a personal device (3) according to the user's selection.
[0041] According to one embodiment of the present invention, an image processing unit (110) can generate a encrypted photon-restricted image by sequentially applying multispectral photon counting imaging and dual random phase encoding to an original image stored in a cloud (1) or a personal device (3) according to the user's selection.
[0042] Additionally, the image processing unit (110) can transmit at least one of the generated photon-limited image, photon-limited encrypted image, and encrypted photon-limited image to the classification unit (130).
[0043] Next, the learning unit (120) can train the first to third classification models to classify images using a plurality of photon-limited images, photon-limited encrypted images, and encrypted photon-limited images. Here, the method by which the learning unit (120) trains the first to third classification models will be described in detail later using FIGS. 3 to 10.
[0044] Specifically, the learning unit (120) can receive initial weights of a classification model from a central server (2), use them to train a first classification model through federated learning to classify images for multiple photon-limited images, and then transmit the weights of the trained first classification model to the central server (2).
[0045] In addition, the central server (2) can calculate a new weight by integrating the weights of multiple classification models and update the weights of the classification models with the calculated new weight.
[0046] Additionally, the learning unit (120) can learn to classify images of a first dataset composed of photon-restricted cipher images or cipher photon-restricted images using adaptive transfer learning, and can learn a second classification model by re-learning to classify images of a second dataset. At this time, the first dataset is a large-scale dataset, and the second dataset is a dataset composed of target data to be classified.
[0047] According to one embodiment of the present invention, the learning unit (120) can train a 2-1 classification model using 1-1 and 2-1 datasets composed of photon-limited cryptographic images.
[0048] According to one embodiment of the present invention, the learning unit (120) can train a 2-2 classification model using 1-2 and 2-2 datasets composed of encrypted photon-restricted images.
[0049] Additionally, the learning unit (120) can train a teacher model to classify images of a second dataset consisting of photon-limited cipher images or cipher photon-limited images using knowledge distillation (KD), and train a student model using the output of the second dataset and the teacher model to train a third classification model.
[0050] According to one embodiment of the present invention, a 3-1 classification model can be trained using a 2-1 dataset composed of photon-limited cryptographic images.
[0051] According to one embodiment of the present invention, a 3-2 classification model can be trained using a 2-2 dataset composed of cryptographic photon-limited images.
[0052] Next, the classification unit (130) can classify an image by applying at least one of the first to third classification models learned to at least one of the generated photon-limited image, photon-limited cipher image, and cipher photon-limited image.
[0053] Specifically, the classification unit (130) can classify the generated photon-limited images using a first classification model transmitted from the central server (2) for the photon-limited images.
[0054] Additionally, the classification unit (130) can classify the photon-restricted cipher image or the cipher-restricted photon image by applying a previously learned second classification model or a third classification model to the generated photon-restricted cipher image or the cipher-restricted photon image.
[0055] Below, a method for classifying images for protecting personal privacy will be explained in more detail using FIGS. 3 to 10.
[0056] FIG. 3 is a flowchart of an image classification method for protecting personal privacy according to another embodiment of the present invention.
[0057] As illustrated in FIG. 3, the image processing unit (110) can generate at least one of a photon-restricted image, a photon-restricted encrypted image, and an encrypted photon-restricted image using a preset method (e.g., Poisson multinomial distribution-based photon counting imaging, multispectral photon counting imaging, dual random phase encoding) according to the user's selection on an original image stored in a cloud (1) or a personal device (3) (S310). Here, the cloud (1) stores an original image (e.g., an open dataset) for pre-training a classification model, and the personal device (3) stores an original image that the user intends to classify.
[0058] Specifically, the image processing unit (110) can generate a photon-limited image using a pre-set method (e.g., photon counting imaging based on Poisson multinomial distribution) on an original image stored in the cloud (1) (S311).
[0059] At this time, the image processing unit (110) can calculate the Poisson multinomial distribution (Poisson(λ(x,y))) using the following [Mathematical Formula 1].
[0060]
[0061] Here, p(xy) is a photon at the pixel in the x-th row and y-th column of the original image (I(x,y)), and λ(x,y) is a previously calculated parameter.
[0062] At this time, the parameter (λ(x,y)) is the expected number of incident photons as shown in [Equation 2] below ( ) and the normalized illuminance of the original image( It is the result of multiplying by ).
[0063]
[0064] Here, h is the total number of pixels in the rows of the original image, and w is the total number of pixels in the columns of the original image.
[0065] In addition, for color images, the image processing unit (110) can simultaneously apply photon-limited images to three channels (R, G, B).
[0066] At this time, the image processing unit (110) can calculate the probability (PMD) of detecting a photon in each pixel of the three channels of the original image through the following [Equation 3].
[0067]
[0068] Here, is the probability of detecting a photon in the R channel, and is the probability of detecting a photon in the G-channel, and is the probability of detecting a photon in the B channel, and is the mean value of the Poisson multinomial distribution in the R channel, and is the mean value of the Poisson multinomial distribution in the G channel, and is the mean value of the Poisson multinomial distribution in channel B.
[0069] Also, the expected number of incident photons ( ) is equal to the sum of the probabilities of detecting photons in each channel.
[0070] Additionally, the image processing unit (110) can generate a photon-limited encrypted image by sequentially applying dual random phase encoding and multispectral photon counting imaging to an original image stored in the cloud (1) or personal device (3) according to the user's selection (S312). At this time, the original image is a color image.
[0071] Specifically, the image processing unit (110) can generate a photon-limited cryptographic image by applying dual random phase encoding to an original image stored in a cloud or personal device (3) and applying multispectral photon counting imaging to the amplitude of the original image to which dual random phase encoding is applied.
[0072] Additionally, the image processing unit (110) can generate a encrypted photon-restricted image by sequentially applying multispectral photon counting imaging and dual random phase encoding to an original image stored in the cloud (1) or personal device (3) according to the user's selection (S313). At this time, the original image is a color image.
[0073] Specifically, the image processing unit (110) can first apply multispectral photon counting imaging to an original image stored in the cloud (1) or personal device (3), and then apply dual random phase encoding to the original image to which multispectral photon counting imaging has been applied to generate a encrypted photon-restricted image.
[0074] Next, the learning unit (120) can train a classification model to classify images for multiple photon-limited images (S320).
[0075] Below, a method for the learning unit (120) to train the first to third classification models using FIGS. 4 to 6 is explained.
[0076] First, a method (S321) for the learning unit (120) to train the first classification model using FIG. 4 is explained.
[0077] In the embodiments described below, the image classification system (100) for protecting personal privacy is specifically described by generating a photon-limited image for an image stored in a personal device (3) and classifying it through a first classification model transmitted from a central server (2). However, the present invention is not limited thereto and can be applied to hosting such as a server or web hosting, and can be applied to data (e.g., documents, images, etc.) containing personal privacy (e.g., the original author's personal information, etc.) stored in the hosting.
[0078] FIG. 4 is a schematic diagram illustrating an image classification method using a first classification model according to another embodiment of the present invention.
[0079] As illustrated in FIG. 4, the central server (2) can initialize the weights of the global model using a plurality of photon-limited images generated by the image processing unit (110) from the first dataset (Original images of general dataset) (① of FIG. 4).
[0080] In addition, the central server (2) has initial weights of the first classification model before the learning unit (120) performs learning ( = ) can be transmitted to individual devices (3) (Client 1 to Client N) respectively (② in FIG. 4). At this time, g represents the weight of the central server (2), k is the index number of the first classification model, and t is the number of repetitions, in the case of the initial weight, t is 0, and the individual devices (3) may be multiple devices.
[0081] In addition, the learning unit (120) receives the initial weights transmitted from the central server (2). ) can be used to train the first classification model by repeating it for the number of local epochs (③ in Fig. 4).
[0082] Here, the learning unit (120) trains a first classification model using initial weights transmitted from the central server (2) on a photon-limited image generated by the image processing unit (110) from a second dataset (Original images of client k's target dataset), and the first classification model can be transmitted from the central server (2) to a personal device (3).
[0083] In addition, the learning unit (120) can update the weights of the first classification model using the learning rate and loss of the local epoch through the following [Equation 4] (④ in FIG. 4).
[0084]
[0085] Here, η is the learning rate of the local epoch, and L is the loss of the local epoch.
[0086] In addition, the learning unit (120) can perform an update by transmitting the weights of the learned first classification model to the central server (2) (⑤ of FIG. 4).
[0087]
[0088] Here, is the total number of photon-limited images for the k-th first classification model, N is the total number of photon-limited images for all first classification models, and k is the total number of first classification models.
[0089] In other words, the central server (2) can train a global model using photon-limited images generated from multiple original images stored in the cloud (1), and the learning unit (120) can train a local model (first classification model) using photon-limited images generated from multiple original images stored in the personal device (3) and initial weights.
[0090] That is, the learning unit (120) can receive the initial weights of the first classification model from the central server (2), use them to train the first classification model through federated learning to classify images for multiple photon-limited images, and then transmit the weights of the trained first classification model to the central server (2).
[0091] As a result, personal data of users is not outsourced during the step of training the classification model (S320) and the classification step (S330), so personal information can be safely protected.
[0092] Next, a method (S322) for the learning unit (120) to train a second classification model using FIG. 5 is described.
[0093] FIG. 5 is a schematic diagram illustrating an image classification method using a second classification model according to another embodiment of the present invention.
[0094] As illustrated in FIG. 5(a), the learning unit (120) can generate a pre-trained model by performing deep learning to classify photon-limited cipher images generated from the first-1 dataset. At this time, a neural network model may be used for deep learning, but is not necessarily limited thereto.
[0095] Additionally, the learning unit (120) can fine-tun the learning model using the photon-limited cryptographic image generated from the 2-1 dataset and generate the 2-1 classification model (Fine-tuned model).
[0096] As illustrated in FIG. 5(b), the learning unit (120) can generate a learning model by performing deep learning to classify cryptographic photon-restricted images generated from the first-second dataset. At this time, a neural network model may be used for deep learning, but is not necessarily limited thereto.
[0097] Additionally, the learning unit (120) can fine-tune the learning model using the cryptographic photon restriction image generated from the second-2 dataset and generate the second-2 classification model (Fine-tuned model).
[0098] Through this, the learning unit (120) can derive a second classification model with improved safety compared to the first classification model.
[0099] Next, a method (S323) for the learning unit (120) to train a third classification model using FIG. 6 is described.
[0100] FIG. 6 is a schematic diagram illustrating an image classification method using a third classification model according to another embodiment of the present invention.
[0101] As illustrated in FIG. 6(a), the learning unit (120) can train a teacher model to classify photon-limited cryptographic images generated from the second-1 dataset (Target dataset).
[0102] Additionally, the learning unit (120) can train a student model using the photon-limited cipher image generated from the 2-1 dataset, the output of the trained teacher model (soft label), and the actual correct answer (ground truth, hard label). At this time, the learning unit (120) can train a 3-1 classification model by simultaneously performing the process of calculating the difference (loss) between the output of the student model (prediction) and the output of the teacher model using the output of the teacher model, and the process of calculating the difference between the output of the student model and the actual correct answer.
[0103] As illustrated in FIG. 6(b), the learning unit (120) can train a teacher model to classify photon-limited cryptographic images generated from the second-2 dataset.
[0104] Additionally, the learning unit (120) can train the student model using the photon-limited cipher image generated from the 2-2 dataset, the output of the trained teacher model, and the actual correct answer. At this time, the learning unit (120) can train the 3-2 classification model by simultaneously performing the process of calculating the difference between the output of the student model and the output of the teacher model using the output of the teacher model, and the process of calculating the difference between the output of the student model and the actual correct answer.
[0105] Through this, the learning unit (120) can derive a third classification model with improved memory efficiency compared to the second classification model.
[0106] Next, the classification unit (130) can classify an image by applying at least one of the first to third classification models learned to at least one of the generated photon-limited image, photon-limited cipher image, and cipher photon-limited image (S330).
[0107] Specifically, the classification unit (130) can classify the image according to the output class (e.g., person, car, bridge, etc.) by applying a learned first classification model to the generated photon-limited image (S331).
[0108] FIG. 7 is a diagram illustrating the classification accuracy of a first classification model according to another embodiment of the present invention.
[0109] Referring to the classification accuracy according to the expected number of incident photons in Fig. 7, it can be seen that the first classification model classifies photon-limited images similarly to classifying original images as the number of incident photons increases.
[0110] Figure 8 is a graph of the compression ratio according to the expected number of incident photons.
[0111] Referring to Fig. 8, it can be seen that the higher the expected number of incident photons, the lower the compression rate of the photon-limited images becomes, similar to the original image, thereby reducing the information loss of the photon-limited images.
[0112] Figure 9 is a graph of the size reduction rate according to the expected number of incident photons.
[0113] Referring to Fig. 9, it can be seen that the higher the expected number of incident photons, the lighter the photon-limited images can be compared to the original images while maintaining the same level of quality.
[0114] Additionally, the classification unit (130) can classify images according to the output class (e.g., person, car, bridge, etc.) by applying a second classification model learned on the generated photon-limited cipher image or the cipher photon-limited image (S332).
[0115] Additionally, the classification unit (130) can classify images according to the output class (e.g., person, car, bridge, etc.) by applying a third classification model learned on the generated photon-limited cipher image or cipher photon-limited cipher image (S333).
[0116] Figure 10 is a graph of classification accuracy according to the expected number of particles of an image encrypted with double random phase encoding and a photon-limited encrypted image according to another embodiment of the present invention.
[0117] Referring to Fig. 10, it can be seen that the second and third classification models classify more accurately even in an environment where photons are limited.
[0118] According to the embodiment of the present invention described above, large-capacity data can be encrypted at a high speed by using an optical-based image processing and encryption technique capable of parallel processing.
[0119] In addition, the security of protecting personal information can be enhanced by combining photon counting with digit-based random phase encoding.
[0120] In addition, by using a lightweight model, the technology can be utilized on devices with high memory efficiency as well as in the cloud.
[0121] In addition, it can prevent attackers or cloud service providers from accessing the original video and save storage space.
[0122] The present invention has been described with reference to the embodiments illustrated in the drawings, but this is merely illustrative, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible therefrom. Accordingly, the true technical scope of protection of the present invention should be determined by the technical spirit of the following claims.
[0123] [Explanation of the symbol]
[0124] 100: Photon-limited image classification system for personal privacy protection
[0125] 1: Cloud 2: Central Server
[0126] 3: Personal device
[0127] 110: Image processing unit
[0128] 120: Learning Department
[0129] 130: Classification section
Claims
1. An image processing unit that generates at least one of a photon-limited image, a photon-limited encrypted image, and an encrypted photon-limited image using a preset method according to the user's selection on an original image stored in a cloud or personal device; and A classification unit that classifies an image using at least one of a first to third classification model learned on at least one of the photon-limited image, photon-limited cryptographic image, and cryptographic photon-limited image of the generated personal device, and Original images for pre-training the classification model are stored in the cloud above, and The above personal device is a photon-limited image classification system for personal privacy protection in which the original image to be classified by the user is stored.
2. In Paragraph 1, A photon restriction image classification system for protecting personal privacy, further comprising a learning unit for training a first to third classification model to classify images using a plurality of photon restriction images, photon restriction encrypted images, and encrypted photon restriction images.
3. In Paragraph 1, The above image processing unit is, A photon-limited image classification system for protecting personal privacy that generates a photon-limited image by applying Poisson multinomial distribution-based photon counting imaging (PMD-PCI) to an original image stored in the cloud or personal device.
4. In Paragraph 3, The above learning unit is, A photon-limited image classification system for protecting personal privacy, which uses initial weights transmitted from a central server, repeatedly trains a first classification model for a number of local epochs, updates the weights of the first classification model using the learning rate and loss of the local epochs, and transmits the trained weights of the first classification model to the central server to perform weight updates.
5. In Paragraph 4, The above classification unit is, A photon restriction image classification system for personal privacy protection that classifies images according to the output class by applying the first classification model learned to the photon restriction image of the generated personal device.
6. In Paragraph 1, The above image processing unit is, A photon-limited image classification system for personal privacy protection that generates a photon-limited encrypted image by sequentially applying Double Random Phase Encoding (DRPE) and Multispectral Photon Counting Imaging (MPCI) to an original image stored in the cloud or personal device.
7. In Paragraph 1, The above image processing unit is, A photon restriction image classification system for personal privacy protection that generates a encrypted photon restriction image by sequentially applying multispectral photon counting imaging and dual random phase encoding to an original image stored in the cloud or personal device.
8. In Paragraph 1, The above learning unit is, Training is performed to classify images of a first dataset composed of photon-limited cipher images or cipher photon-limited images using adaptive transfer learning, and retraining is performed to classify images of a second dataset composed of photon-limited cipher images or cipher photon-limited images to train the second classification model. The first dataset mentioned above is a large dataset, and The above second dataset is a photon-limited image classification system for personal privacy protection, which is a dataset composed of target data to be classified.
9. In Paragraph 8, The above classification unit is, A photon restriction image classification system for personal privacy protection that classifies images according to the output class by applying the second classification model learned to the photon restriction encrypted image or the encrypted photon restriction image of the generated personal device.
10. In Paragraph 1, The above learning unit is, A teacher model is trained to classify images of a second dataset consisting of photon-limited cryptographic images or cryptographic photon-limited images using knowledge distillation (KD), and a third classification model is trained by training a student model using the second dataset and the output of the teacher model. The above second dataset is a photon-limited image classification system for personal privacy protection, which is a dataset composed of target data to be classified.
11. In Paragraph 10, The above classification unit is, A photon restriction image classification system for personal privacy protection that classifies images according to the output class by applying the third classification model learned to the photon restriction encrypted image or the encrypted photon restriction image of the generated personal device.
12. A step in which an image processing unit generates at least one of a photon-limited image, a photon-limited encrypted image, and an encrypted photon-limited image using a preset method according to the user's selection on an original image stored in a cloud or personal device; and The method includes the step of classifying an image using at least one of a first to third classification model trained on at least one of the photon-limited image, photon-limited cipher image, and cipher photon-limited image of the personal device in which the classification unit is generated. Original images for pre-training the classification model are stored in the cloud above, and The above personal device is a photon-limited image classification method for protecting personal privacy in which the original image to be classified by the user is stored.
13. In Paragraph 12, A photon restriction image classification method for protecting personal privacy, further comprising the step of a learning unit training a first to third classification model to classify images using a plurality of photon restriction images, photon restriction cryptographic images, and cryptographic photon restriction images.