CN114925848A - Target detection method based on transverse federated learning framework - Google Patents
Target detection method based on transverse federated learning framework Download PDFInfo
- Publication number
- CN114925848A CN114925848A CN202210492162.6A CN202210492162A CN114925848A CN 114925848 A CN114925848 A CN 114925848A CN 202210492162 A CN202210492162 A CN 202210492162A CN 114925848 A CN114925848 A CN 114925848A
- Authority
- CN
- China
- Prior art keywords
- model
- client
- detection model
- parameters
- learning framework
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/602—Providing cryptographic facilities or services
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
- G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
- G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2221/00—Indexing scheme relating to security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F2221/21—Indexing scheme relating to G06F21/00 and subgroups addressing additional information or applications relating to security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F2221/2107—File encryption
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioethics (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biomedical Technology (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Computer Security & Cryptography (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a target detection method based on a transverse federated learning framework, which comprises the steps of constructing the transverse federated learning framework, wherein the framework comprises a client detection model, a central service network and a data set; allocating an independent data set for each client detection model, and training the model through the data set; optimizing a gradient descent algorithm used in the training process of the client detection model, and accelerating the convergence speed of the model; extracting parameters of the trained model, and transmitting the parameters to a central service network through a differential privacy strategy; training the central service network, obtaining model parameters after parameter aggregation, returning the model parameters to each client detection model, and updating the model by the client detection model according to the returned parameters; and detecting by using the updated model. The invention adopts the target detection method based on the horizontal federal learning framework, and can solve the problems that the existing target detection method has low accuracy and efficiency and the privacy of data can not be protected.
Description
Technical Field
The invention relates to the technical field of target detection, in particular to a target detection method based on a transverse federated learning framework.
Background
With the advent of the information-oriented society, computer vision technology has gradually changed the lifestyle of humans. Common application scenarios include electronic payment, medical image detection, a skylight system, and the like. The large amount of complex visual information data provides a rich sample for the processing and analysis capability of the computer, and also brings a serious challenge. Object detection is one of the important topics in the field of computer vision, and its task is to classify objects of interest and label corresponding locations in an image. In recent years, the development of deep learning provides strong driving force for the field of computer vision, and the deep learning has the characteristics of high-efficiency learning and wide coverage range. Therefore, people gradually apply deep learning to the target detection task and achieve good effect.
In the traditional target detection algorithm, candidate box positioning, feature extraction and classifier judgment are the key points of research. The detection steps can be summarized as follows: firstly, carrying out target positioning on the image by a sliding window method, then carrying out feature extraction by using a feature extraction algorithm, and finally judging by a classifier. Common classifiers are support vector machines, Adaboost, etc. Although the detection process of the conventional method is simple and easy to understand, there are two obvious problems to be solved: 1) the region selection module does not adopt a targeted strategy, which results in the generation of a repeatedly invalid window and a too long region selection time. 2) The features extracted by the feature extraction module are artificially designed, and the robustness of the features is poor, so that the accuracy of classification is directly influenced.
The appearance of deep learning enables the quality of the target detection field to be improved. The target detection algorithm based on deep learning combines the advantages of the classifier and the convolutional neural network, and can enhance the capability of feature extraction and efficient classification of the detection model. The Faster R-CNN is a target detection model which is used more at present and can more efficiently and accurately acquire candidate frames through a region generation network. Moreover, a characteristic diagram pyramid network can be embedded into the model to solve the multi-scale problem in object detection, and the detection effect is greatly improved by changing the connection mode of the network under the condition that the calculation amount of the original model is basically not increased.
In real life, target detection is widely applied to multiple fields such as area monitoring and traffic management. In these applications, data exchange and cooperation among multiple models are often required, so as to form an enhanced regional joint detection network. For example, in the detection of traffic signs, pictures are collected at each detection point for local model training, but the ideal detection effect cannot be ensured due to the small data volume of the collected pictures. In addition, the privacy of the data is not protected during the object detection process.
Disclosure of Invention
The invention aims to provide a target detection method based on a horizontal federated learning framework, and solves the problems that the existing target detection method is low in accuracy and efficiency, and the privacy of data cannot be protected.
In order to achieve the purpose, the invention provides a target detection method based on a horizontal federal learning framework, which comprises the following steps:
s1, constructing a transverse federal learning framework, wherein the framework comprises a client detection model, a central service network and a data set;
s2, distributing independent same source data sets for each client detection model, and training the models through the data sets;
s3, optimizing the training process of the client detection model by using a gradient descent algorithm to accelerate the convergence rate of the model;
s4, extracting parameters of the trained client detection model, and transmitting the parameters to a central service network through a differential privacy strategy;
and S5, training the central service network, obtaining client detection model parameters after parameter aggregation, returning the client detection model parameters to each client detection model, updating the client detection model according to the returned parameters by the client detection model, and detecting by using the updated client detection model.
Preferably, in step S1, a client detection model based on fast R-CNN is selected, and the FPN network is embedded in the model for generating feature maps with different scales.
It is preferable thatIn step S3, in the training process, an Adam algorithm is used for optimization, and the Adam algorithm utilizes the historical gradient mean m of the parameter t Sum parameter historical gradient squared sum mean v t The learning rate is adaptively adjusted, and information can be learned at every moment; the two mean values are calculated as follows:
m t =β 1 m t-1 +(1-β 1 )g t
v t =β 2 v t-1 +(1-β 2 )g t 2
wherein, beta 1 、β 2 Represents the attenuation factor, g t At time t θ t A gradient value of (a);
get beta 1 =0.9,β 2 =0.999,ε=10 -8 Correcting the above formula;
where α is the learning rate.
Preferably, in step S4, the data is encrypted homomorphically by the differential privacy policy to prevent the gradient information from leaking, where x is indicated by [ [ x ] ] indicating that the encryption operation is performed on x, and the homomorphic encryption formula is as follows:
where u and v represent any two elements in the plaintext space, the encryption result is [ [ u ]]]And [ [ v ]]],Then represents an addition or multiplication operation, Dec sk Representing a decryption operation.
Preferably, in step S5, parameter aggregation is implemented based on an average influence degree algorithm, and a weight is assigned to each client according to the influence degree of each client detection model; the number of the data sets is m, the number of the client detection models is n, and the influence factor matrix X and the output Y are respectively expressed as:
Y=[y 1 y 2 …y m ] T
first, the ith influencing factor is fine-tuned by delta, which means that the value is increased or decreased by 10%, and two new groups of inputs and outputs are respectively represented asAnd
secondly, calculating the MIV value of the client, wherein the value represents the relative influence degree of the current client, and the calculation formula is as follows:
finally, calculating the corresponding contribution degree C of each client detection model i And the average impact value C, and,
the target detection method based on the transverse federated learning framework has the advantages and positive effects that:
1. according to the invention, a horizontal federated learning framework is applied to a target detection task, the purpose of regional joint detection is realized, parameters are updated by sharing data among regional models, and model convergence can be achieved more quickly compared with a traditional centralized training method.
2. The invention utilizes the advantage that the FPN network has cross-layer connection, and can have better detection effect on multi-scale images. The FPN network comprises three connection modes: the bottom-up connection relies on forward propagation, and the size of the feature map is usually reduced after convolution operation in the process, so that the global information contained in the deep features is rich. Top-down connection upsamples deep features, thereby enlarging the size of the feature map. The lateral connection is a feature map obtained by combining the two previous features at the same size, and high-resolution information and high-semantic information can be simultaneously enhanced in the feature map.
3. The method introduces an attention mechanism during parameter aggregation, the attention mechanism is realized by adopting an average influence degree (MIV) weighting algorithm, the algorithm is used for evaluating the influence degree of each region model on a combined detection result, and the aggregation effect is improved by strengthening a local model with large influence degree and weakening the local model with small influence degree.
The technical solution of the present invention is further described in detail by the accompanying drawings and embodiments.
Drawings
FIG. 1 is a flowchart of an embodiment of a target detection method based on a horizontal federated learning framework in the present invention;
FIG. 2 is a FPN network structure diagram of an embodiment of a target detection method based on a horizontal federated learning framework according to the present invention;
FIG. 3 is a model after combination of fast R-CNN and FPN networks according to an embodiment of a target detection method based on a horizontal federated learning framework of the present invention;
FIG. 4 is a federated learning framework in accordance with an embodiment of a method for target detection based on a horizontal federated learning framework of the present invention;
FIG. 5 is a graph showing the comparison of the accuracy of the horizontal federated learning training and the centralized training in an embodiment of the target detection method based on the horizontal federated learning framework according to the present invention;
fig. 6 is a loss value comparison diagram of the horizontal federal learning training and the centralized training in the embodiment of the target detection method based on the horizontal federal learning framework in the present invention.
Detailed Description
The technical solution of the present invention is further illustrated by the accompanying drawings and examples.
Examples
Fig. 1 is a flowchart of an embodiment of an object detection method based on a horizontal federal learning framework in the present invention, fig. 2 is a structure diagram of a FPN network in the embodiment of an object detection method based on a horizontal federal learning framework in the present invention, fig. 3 is a model of a combination of fast R-CNN and a FPN network in the embodiment of an object detection method based on a horizontal federal learning framework in the present invention, and fig. 4 is a federal learning framework in the embodiment of an object detection method based on a horizontal federal learning framework in the present invention. As shown in the figure, a target detection method based on a horizontal federal learning framework comprises the following steps:
s1, constructing a horizontal federal learning framework, wherein the framework mainly comprises a client detection model, a central service network and a data set. The data set adopts a PASCAL VOC public data set which is suitable for tasks such as target detection, image segmentation and the like and comprises twenty categories. The fast R-CNN is selected as a detection model of a client, the FPN is embedded into the model to generate feature maps with different scales, and the position information of a shallow network and the semantic information of a deep network are fully combined by fusing the feature maps, so that the multi-scale problem in a target detection task is relieved.
And S2, allocating independent same-source data sets for each client detection model, and training the model through the data sets.
And S3, optimizing the client detection model by using a gradient descent algorithm in the training process, and accelerating the convergence speed of the model. Adam algorithm is used for optimization in the training process, wherein Adam is a first-order optimization algorithm capable of replacing the traditional random gradient descent process, and can iteratively update the weight of the neural network based on training data. The basic mechanism of the Adam optimization algorithm is different from that of the traditional random gradient descent algorithm, and the latter method is to keep a single learning rate to update all weights, and the weights are not changed in the process; the former designs different adaptive learning rates for different parameters independently by calculating the first moment estimate and the second moment estimate of the gradient.
The Adam algorithm not only realizes the self-adaptive adjustment of the learning rate, but also can keep learning information at every moment. This is by means of the historical gradient mean m of the parameter t Sum parameter historical gradient squared sum mean v t And (4) realizing.
m t =β 1 m t-1 +(1-β 1 )g t
v t =β 2 v t-1 +(1-β 2 )g t 2
Wherein, beta 1 、β 2 Represents the attenuation factor, g t At time t θ t The gradient value of (a). At model initialization, when beta 1 And beta 2 When the values are all 1, the attenuation rate is small and the learning rate is 0. To solve this problem, m needs to be matched t And v t The following offset correction is performed, and β is usually taken 1 =0.9,β 2 =0.999,ε=10 -8 The effect is better.
The parameter update equation at this time is:
where α is the learning rate.
And S4, extracting the parameters of the trained client detection model, and transmitting the parameters to the central service network through the differential privacy policy. In order to ensure the safety of data, the central server introduces Gaussian noise to perform differential privacy protection on important information when updating global model parameters. The differential privacy provides guarantee for information theoretical security, and a random mechanism is adopted, so that output distribution can not be changed greatly after a certain sample in input samples is changed, and the external part can not judge which party the input data provides through the output data distribution, thereby being a reliable security mechanism. The data is encrypted in a homomorphic way through a differential privacy strategy, gradient information leakage is prevented, x is encrypted and expressed by [ [ x ] ], and a homomorphic encryption formula is as follows:
where u and v represent any two elements in the plaintext space, the result of the encryption is [ [ u ]]]And [ [ v ]]],Then represents an addition or multiplication operation, Dec sk Representing a decryption operation.
And S5, training the central service network, obtaining client detection model parameters after parameter aggregation, returning the client detection model parameters to each client detection model, updating the client detection model according to the returned parameters by the client detection model, and detecting by using the updated client detection model.
Parameter aggregation is realized based on an average influence degree algorithm, and weight is distributed to each client according to the influence degree of each client detection model. The number of the data sets is m, the number of the client detection models is n, and the influence factor matrix X and the output Y are respectively expressed as:
Y=[y 1 y 2 …y m ] T
first, the ith influencing factor is fine-tuned by delta, which means that the value is increased or decreased by 10%, and two new groups of inputs and outputs are respectively represented asAnd
secondly, calculating an MIV value of the client detection model, wherein the value represents the relative influence degree of the current client detection model, and the calculation formula is as follows:
finally, calculating the corresponding contribution degree C of each client detection model i And the average impact value C, and,
the central server training steps are as follows:
1) let communication round t be 0 and start federal learning training.
2) The privacy loss delta of the current round is calculated by using an MA (moment account) mechanism of the server, and the calculation formula is as follows:
δ←MA(σ t ,ε,K)
wherein σ t Representing Gaussian noise of the t-th round, epsilon representing a privacy bearing value, K representing the number of client detection models, stopping the training of the round and returning if delta is larger than QOtherwise proceed to the next step, Q refers to the loss threshold.
3) The global model parameters of the current roundAnd an adaptive training period tau t And broadcasting and sending the broadcast to K client detection models. Tau is t The calculation formula of (2) is as follows:
wherein, tau t Representing a local training period, τ, of the t-th round of communication 0 Representing the initial local training period, η 0 ,η t Respectively, the initial learning rate and the learning rate of the t-th communication, l represents the objective function of the model,global model parameters representing the t-th communication.
4) The central service network receives all client detection model parameters to obtain parameter update valuesAnd detecting the median M of the model parameter updating norm through the client, and simultaneously calculating the Gaussian noise parameter S of the turn.
5) Calculating the mean value X (i, j) and the variance Y (i, j) of the (i, j) th parameter in the updated parameters of the client detection model in the current round 2 And the variance σ of all update parameters of the client detection model 2 。
Wherein, the gaussian noise formula is:
6) and calculating global model parameters, updating the communication round T to be T +1, jumping to the step 2 if T is less than the total communication round number T, and returning to a current value if not. The global model parameter calculation formula is as follows:
fig. 5 is a comparison diagram of accuracy of the horizontal federal learning training and the centralized training in an embodiment of the target detection method based on the horizontal federal learning frame of the present invention, and fig. 6 is a comparison diagram of loss values of the horizontal federal learning training and the centralized training in an embodiment of the target detection method based on the horizontal federal learning frame of the present invention. As shown in the figure, finally, an experiment is performed by using the updated model, the experiment includes 10 clients, 5 of the clients are randomly selected to participate in training in each round, each client needs to perform iterative training 3 times to detect the model, and the number of global communication rounds is set to 25 times. The centralized training is equivalent to only one client participating in the training in each round, and the local iteration times are respectively set to 1, 2 and 3 for comparison in order to obtain high-performance configuration. As can be seen from the accompanying fig. 5 and fig. 6 in the specification, the target detection method adopting the horizontal federal learning framework has higher accuracy and lower loss value.
Therefore, the invention adopts the target detection method based on the horizontal federal learning framework, and can solve the problems that the existing target detection method has low accuracy and efficiency and the privacy of data cannot be protected.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the invention without departing from the spirit and scope of the invention.
Claims (5)
1. A target detection method based on a horizontal federal learning framework is characterized by comprising the following steps:
s1, constructing a transverse federated learning framework, wherein the framework comprises a client detection model, a central service network and a data set;
s2, distributing independent same-source data sets for each client detection model, and training the models through the data sets;
s3, optimizing the client detection model by using a gradient descent algorithm in the training process, and accelerating the convergence speed of the model;
s4, extracting parameters of the trained client detection model, and transmitting the parameters to a central service network through a differential privacy strategy;
and S5, training the central service network, obtaining client detection model parameters after parameter aggregation, returning the client detection model parameters to each client detection model, updating the client detection model according to the returned parameters by the client detection model, and detecting by using the updated client detection model.
2. The method for target detection based on the horizontal federated learning framework as claimed in claim 1, wherein: in step S1, a client detection model based on fast R-CNN is selected, and the FPN network is embedded in the model for generating feature maps of different scales.
3. The method for target detection based on the horizontal federated learning framework as claimed in claim 1, wherein: in the step S3, in the training process, Adam algorithm is used for optimization, and the Adam algorithm utilizes the historical gradient mean value m of the parameter t Sum parameter historical gradient squared sum mean v t The learning rate is adaptively adjusted, and information can be learned at every moment; the two mean values are calculated as follows:
m t =β 1 m t-1 +(1-β 1 )g t
v t =β 2 v t-1 +(1-β 2 )g t 2
wherein, beta 1 、β 2 Represents the attenuation factor, g t At time t θ t A gradient value of (a);
get beta 1 =0.9,β 2 =0.999,ε=10 -8 Correcting the above formula;
where α is the learning rate.
4. The method for target detection based on the horizontal federated learning framework as claimed in claim 1, wherein: in step S4, the data is homomorphic encrypted by the differential privacy policy to prevent the gradient information from leaking, where x is denoted by [ [ x ] ] and the homomorphic encryption formula is as follows:
5. The method for target detection based on the horizontal federated learning framework as claimed in claim 1, wherein: in step S5, parameter aggregation is implemented based on an average influence algorithm, and a weight is assigned to each client according to the influence of the detection model; the number of the data sets is m, the number of the client detection models is n, and the influence factor matrix X and the output Y are respectively expressed as:
Y=[y 1 y 2 …y m ] T
first, the ith influencing factor is trimmed by δ, and two new sets of inputs and outputs are represented asAnd Y i ± 。
Secondly, calculating the MIV value of the client, wherein the value represents the relative influence degree of the current client, and the calculation formula is as follows:
finally, calculating the corresponding contribution degree C of each client detection model i And the average impact value C, and,
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210492162.6A CN114925848A (en) | 2022-05-07 | 2022-05-07 | Target detection method based on transverse federated learning framework |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210492162.6A CN114925848A (en) | 2022-05-07 | 2022-05-07 | Target detection method based on transverse federated learning framework |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114925848A true CN114925848A (en) | 2022-08-19 |
Family
ID=82805927
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210492162.6A Pending CN114925848A (en) | 2022-05-07 | 2022-05-07 | Target detection method based on transverse federated learning framework |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114925848A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117576655A (en) * | 2023-11-04 | 2024-02-20 | 河南工业大学 | Traffic sign detection method and system based on federal learning |
-
2022
- 2022-05-07 CN CN202210492162.6A patent/CN114925848A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117576655A (en) * | 2023-11-04 | 2024-02-20 | 河南工业大学 | Traffic sign detection method and system based on federal learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113688723B (en) | Infrared image pedestrian target detection method based on improved YOLOv5 | |
CN112668726A (en) | Personalized federal learning method with efficient communication and privacy protection | |
CN103700114B (en) | A kind of complex background modeling method based on variable Gaussian mixture number | |
CN114358912A (en) | Risk weight fusion anomaly detection method based on federal learning | |
US20210241173A1 (en) | Device-free localization robust to environmental changes | |
CN117150566B (en) | Robust training method and device for collaborative learning | |
CN114925848A (en) | Target detection method based on transverse federated learning framework | |
Pang et al. | Federated learning for crowd counting in smart surveillance systems | |
CN115310625A (en) | Longitudinal federated learning reasoning attack defense method | |
CN115422537A (en) | Method for resisting turnover attack of federal learning label | |
CN114582011A (en) | Pedestrian tracking method based on federal learning and edge calculation | |
CN116644439B (en) | Model safety assessment method based on denoising diffusion model | |
CN116186629B (en) | Financial customer classification and prediction method and device based on personalized federal learning | |
Racherla et al. | Deep-IDS: A Real-Time Intrusion Detector for IoT Nodes Using Deep Learning | |
CN116259057A (en) | Method for solving data heterogeneity problem in federal learning based on alliance game | |
CN115879108A (en) | Federal learning model attack defense method based on neural network feature extraction | |
CN112749686B (en) | Image detection method, image detection device, computer equipment and storage medium | |
CN115630361A (en) | Attention distillation-based federal learning backdoor defense method | |
CN113159317B (en) | Antagonistic sample generation method based on dynamic residual corrosion | |
CN116129534A (en) | Image living body detection method and device, storage medium and electronic equipment | |
Lin et al. | Collaborative Framework of Accelerating Reinforcement Learning Training with Supervised Learning Based on Edge Computing | |
Zhang et al. | Owl: A pre-and post-processing framework for video analytics in low-light surroundings | |
CN112862001B (en) | Privacy protection method and system for decentralizing data modeling under federal learning | |
CN117176923B (en) | Intelligent community police service patrol method and system based on data encryption | |
US20230237315A1 (en) | Method and system for efficient learning on large multiplex networks |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |