CN118379608B - High-robustness deep forgery detection method based on self-adaptive learning - Google Patents
High-robustness deep forgery detection method based on self-adaptive learning Download PDFInfo
- Publication number
- CN118379608B CN118379608B CN202410834314.5A CN202410834314A CN118379608B CN 118379608 B CN118379608 B CN 118379608B CN 202410834314 A CN202410834314 A CN 202410834314A CN 118379608 B CN118379608 B CN 118379608B
- Authority
- CN
- China
- Prior art keywords
- face image
- quality
- algorithm
- degradation
- adaptive
- 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.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 86
- 238000006731 degradation reaction Methods 0.000 claims abstract description 83
- 230000015556 catabolic process Effects 0.000 claims abstract description 78
- 230000003044 adaptive effect Effects 0.000 claims abstract description 34
- 238000012549 training Methods 0.000 claims abstract description 33
- 238000005070 sampling Methods 0.000 claims abstract description 27
- 238000000034 method Methods 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 15
- 238000007906 compression Methods 0.000 claims description 9
- 238000013441 quality evaluation Methods 0.000 claims description 8
- 230000006835 compression Effects 0.000 claims description 7
- 230000009977 dual effect Effects 0.000 claims description 6
- 238000009826 distribution Methods 0.000 claims description 5
- 238000005242 forging Methods 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000005457 optimization Methods 0.000 claims description 3
- 238000001303 quality assessment method Methods 0.000 claims description 2
- 239000000284 extract Substances 0.000 claims 1
- 230000004580 weight loss Effects 0.000 description 8
- 230000006872 improvement Effects 0.000 description 5
- 230000000593 degrading effect Effects 0.000 description 4
- 230000005540 biological transmission Effects 0.000 description 3
- 238000012512 characterization method Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 2
- 230000000750 progressive effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000007480 spreading Effects 0.000 description 1
- 238000003892 spreading Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/95—Pattern authentication; Markers therefor; Forgery detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Databases & Information Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a high-robustness deep counterfeiting detection method based on self-adaptive learning, which is characterized in that firstly, degraded face images with different qualities are self-adaptively generated based on a quality degradation generation algorithm, the quality diversity of the existing deep counterfeiting face image dataset is supplemented, meanwhile, learning signals of the face images with different qualities are coordinated by combining an adaptive sampling network, and counterfeiting characteristics of face images with unknown quality are dynamically captured, so that the detection performance of a deep counterfeiting detection model on low-quality counterfeiting images is improved. The invention helps the self-adaptive depth fake detection algorithm to better learn the difference between fake features and degradation noise, improves the robustness of the self-adaptive depth fake detection algorithm to the face image with unknown quality, carries out algorithm dynamic training based on the face image quality, guides the self-adaptive depth fake detection algorithm to optimize, and improves the balance between the accuracy of the self-adaptive depth fake detection algorithm to the original face image and the degradation face image.
Description
Technical Field
The invention relates to the field of image classification based on Artificial Intelligence (AI), in particular to a high-robustness deep forgery detection method based on self-adaptive learning.
Background
With the development of deep forging technology, the deep forging technology becomes a tool for spreading false contents by malicious users due to the convenience and authenticity of forging face images. In order to prevent misuse of face depth forgery technology by malicious users, a depth forgery detection algorithm aims to identify whether an input face image is a depth forgery face image.
Most of the existing depth counterfeiting detection algorithms aim at high-quality counterfeiting face images, however, in order to ensure transmission efficiency, degradation operations of reducing the quality of the face images, such as compression, clipping, downsampling and the like, are carried out in the transmission process, and the degradation operations of the quality of the damaged face images can effectively improve the transmission speed, but fine noise is introduced into the face images at the same time to interfere with identifiable counterfeiting features in the counterfeiting face images, so that the existing depth counterfeiting detection algorithms cannot identify the degraded depth counterfeiting face images. The invention researches the influence of the degradation operation of the face image suffered by the depth fake image in the propagation process on the robustness of the depth fake detection, and focuses on various unknown and complex quality degradation operations of the depth fake face image on different platforms. Research results show that when the existing deep counterfeiting detection algorithm detects a fake face image, if the face image is subjected to the face image degradation operation, the extracted face features are very easy to mix degradation noise when the feature of the face image is extracted by the deep counterfeiting detection algorithm, so that the algorithm performance is obviously reduced when the low-quality deep counterfeiting face image is detected by the deep counterfeiting detection algorithm. Therefore, a depth forgery detection method for high-robustness progressive adaptive face image quality of degraded images is needed to improve the detection performance of the depth forgery detection model on low-quality images.
The invention has the following technical problems:
1) The effectiveness of existing depth forgery detection algorithms depends on their ability to discern nuances between genuine and fake face images. However, the face image quality degradation operation will introduce noise similar to the counterfeit feature to interfere with discrimination of minor differences between genres and genres, ultimately affecting the effectiveness of the deep counterfeit detection algorithm;
2) The disclosed depth fake face image data set adopted by the existing depth fake detection algorithm cannot cover various face image quality degradation, and although the data enhancement technology used by the existing algorithm in the training process can alleviate the problem, the diversity of data enhancement is obviously different from the real-world complex and diverse degradation operation, so that the accuracy of the existing depth fake detection algorithm on the degraded face image is poor;
3) The existing depth fake detection algorithm adopts a random sampling method to train a model, so that inconsistent feedback of face images from different qualities is difficult to balance, and the depth fake detection algorithm cannot adapt to quality changes of different face images.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a high-robustness deep counterfeiting detection method based on self-adaptive learning. According to the method, degraded face images with different qualities are adaptively generated based on a quality degradation generation algorithm, quality diversity of an existing deep fake face image data set is supplemented, learning signals of the face images with different qualities are coordinated by combining an adaptive sampling network, fake characteristics of face images with unknown quality are dynamically captured, and therefore detection performance of a deep fake detection model on low-quality fake images is improved.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
the invention discloses a high-robustness deep forgery detection method based on self-adaptive learning, which comprises the following steps: an algorithm training phase and an algorithm reasoning phase, in particular:
Algorithm training phase:
Obtaining an original face image Corresponding label；
Generating an algorithm by quality degradationGenerating a degradation setting；
Quality degradation generation algorithmAccording to degradation settingsAnd the obtained original face imageGenerating an original face imageCorresponding degraded face imageObtaining a human face image pairCorresponding real tag pairs；
Based on the face image pairAdaptive depth forgery detection algorithmPerforming dual tasks of image quality evaluation and depth forgery detection to obtain corresponding original face images respectivelyAnd degrading the face imageIs a pair of outputs of (a) and (b): image quality tag pair for evaluating image qualityAnd detecting label pair for judging image authenticity；
Tag pairs based on image qualityAnd detecting tag pairsAdaptive sampling networkFace image quality feedback obtained through calculationAnd algorithm performance feedback；
Adaptive sampling networkFeedback according to algorithm performanceAnd face image quality feedbackComputing pairs of images relating to facesTraining weights of (a)Training weightsIs used to calculate the weight loss and finally the weight loss is used to assist the quality degradation generation algorithmAnd adaptive depth forgery detection algorithmUpdating algorithm parameters;
algorithm reasoning:
Obtaining a face image to be detected ；
Inputting the data to a trained adaptive depth forgery detection algorithmAdaptive depth forgery detection algorithmCalculating to obtain classification result of deep forgery detection。
As a further improvement, the invention relates to a generation algorithm by quality degradationGenerating a degradation settingThe method specifically comprises the following steps:
Quality degradation generation algorithm Four classical degradation algorithms of Gaussian noise, gaussian blur, downsampling and JPEG compression are adopted to simulate degradation operation suffered by an image in the propagation process and set in degradationMiddle constraint Gaussian blur kernelDownsampling sizeGaussian noise sizeJPEG compression ratioIs used for the value range of the (a), quality degradation generation algorithmIs a backbone network of (1)Mainly for predicting degradation settingsIn (a)、、、The value of (2) is as follows:
。
as a further improvement, the quality degradation generation algorithm of the invention According to degradation settingsAnd the obtained original face imageGenerating an original face imageCorresponding degraded face imageObtaining a human face image pairThe method specifically comprises the following steps:
Quality degradation generation algorithm According to the backbone networkPredicted degradation settingsFor the original face imagePerforming degradation operation to generate a degraded face image：
；
Wherein, A convolution operation is represented and is performed,Representing scale factorsIs used for the down-sampling operation of (a),Representation ofCompression ratio.
As a further improvement, the invention relates to the face image pairAdaptive depth forgery detection algorithmPerforming dual tasks of image quality evaluation and depth forgery detection to obtain corresponding original face images respectivelyAnd degrading the face imageIs a pair of outputs of (a) and (b): image quality tag for evaluating image qualityAnd detecting label pair for judging image authenticityThe method specifically comprises the following steps:
adaptive depth forgery detection algorithm From the original face images respectivelyAnd corresponding degraded face imageExtraction of image representationsAndCharacterization input to adaptive depth counterfeit detection algorithmIs a backbone network of (1)Quality tag pair for evaluating image qualityAnd detecting label pair for judging image authenticityAccording to quality label pairsAdaptive depth forgery detection algorithmIs described which introduces a quality order loss function into the image quality assessment task:
；
Wherein, The number of face images is represented,AndRespectively the firstZhang Yuanshi predictive quality labels for face images and degraded face images,Is a super parameter, and the depth falsification detection task loss function can be expressed as:
；
Wherein, Represent the firstThe real label of the face image is opened,Represent the firstZhang Yuanshi the prediction probability of the fake face image corresponding to the face image.
As a further improvement, the label pair according to the image quality of the inventionAnd detecting label pair for judging image authenticityAdaptive sampling networkFace image quality feedback obtained through calculationAnd algorithm performance feedbackThe method specifically comprises the following steps:
adaptive sampling network based on image quality tags Computing image quality feedbackTo represent a backbone networkThe higher the predicted image quality score, the better the image quality:
；
Wherein, Representing gaussian distribution and adaptive sampling networkCalculating a real tag from a detected tagPredictive tagsAlgorithm performance feedback between。
As a further improvement, the training weight of the inventionIs used to calculate the weight loss and finally the weight loss is used to assist the quality degradation generation algorithmAnd adaptive depth forgery detection algorithmUpdating algorithm parameters, specifically:
based on self-adaptive sampling network Is the original face imageAnd corresponding degraded face imagesAdaptively generated corresponding weightsAndQuality degradation generation algorithmAnd adaptive depth forgery detection algorithmThe training loss function of (2) is as follows:
，
，
Wherein, Representing training face image weights, including raw face imagesAnd corresponding degraded face images，Non-negative super parameters for the balance loss function;
adaptive sampling network Predicting weight of each face image in current training iteration by adopting local optimal strategy, and self-adapting sampling networkThe optimization function of (2) is as follows:
，
Wherein, Is a super parameter.
The beneficial effects of the invention are as follows:
1) The invention generates the algorithm through quality degradation Generating a degradation settingThe degradation arrangementFour classical degradation algorithms of Gaussian noise, gaussian blur, downsampling and JPEG compression are adopted to generate an algorithm for quality degradationVarious degradation modes are provided to accurately fit degradation operations on images in a real scene including WeChat, microblog, facebook, twitter and other platforms.
2) The quality degradation generation algorithm of the inventionAccording to degradation settingsAnd the obtained original face imageGenerating an original face imageCorresponding degraded face imageObtaining a human face image pair. Quality degradation generation algorithmAccording to degradation settingsAnd original face imageAdaptively generating degraded face images of different qualitiesThe problem that the existing disclosed deep fake face image dataset cannot cover the quality degradation of various face images is solved, the quality diversity of the existing deep fake face image dataset is supplemented, the self-adaptive deep fake detection algorithm is helped to learn the difference between fake features and degradation noise better, and assistance is provided for the self-adaptive deep fake detection algorithm to effectively capture the fake features;
3) The invention relates to a human face image pair Adaptive depth forgery detection algorithmPerforming dual tasks of image quality evaluation and depth forgery detection to obtain corresponding original face images respectivelyAnd degrading the face imageIs a pair of outputs of (a) and (b): face image quality label pair for evaluating face image qualityAnd a detection label pair for judging the authenticity of the deeply forged face image. The self-adaptive depth counterfeiting detection algorithm is used for distinguishing the characteristic difference between the original face image and the degraded face image based on contrast learning, effectively describing the characteristic distribution of the face images with different quality, executing the dual tasks of face image quality evaluation and depth counterfeiting detection, and improving the robustness of the self-adaptive depth counterfeiting detection algorithm to the face images with unknown quality;
4) The invention relates to a self-adaptive sampling network Feedback according to human face image qualityAnd algorithm performance feedbackComputing pairs of images relating to facesTraining weights of (a)The training weightIs used to calculate a weight loss, and finally the weight loss is used to assist the quality degradation generation algorithmAnd adaptive depth forgery detection algorithmUpdating algorithm parameters. The invention adopts progressive weighting and optimizing method, and feeds back according to the quality of training imageAnd algorithm performance feedbackThe corresponding weight of the training image is gradually adjusted so as to realize the self-adaptive depth counterfeiting detection of the face image quality, the algorithm dynamic training is carried out based on the face image quality, the self-adaptive depth counterfeiting detection algorithm is guided to optimize according to inconsistent feedback generated by the face images with different qualities, and the balance between the accuracy of the self-adaptive depth counterfeiting detection algorithm on the original face image and the degraded face image is improved.
Drawings
FIG. 1 is a training schematic diagram of a high robustness deep forgery detection method based on adaptive learning in the present invention;
Fig. 2 is a schematic diagram of reasoning of the high robustness deep forgery detection method based on adaptive learning in the present invention.
Detailed Description
In order to facilitate an understanding and practice of the invention by those of ordinary skill in the art, the invention will be described in further detail below with reference to the drawings and specific examples, it being understood that the examples described herein are for illustration and explanation only and are not intended to be limiting of the invention.
Aiming at the problems that the existing depth forgery detection method has poor robustness and cannot adapt to the quality change of a face image and finally the accuracy of a degraded face image is reduced, the invention provides a high-robustness depth forgery detection method based on self-adaptive learning, and fig. 1 is a training schematic diagram of the high-robustness depth forgery detection method based on self-adaptive learning;
the specific implementation method of the invention is as follows:
Algorithm training phase:
Step one: obtaining an original face image Corresponding label；
Step two: quality degradation generation algorithmGenerating a degradation setting. Quality degradation generation algorithmSetting degradation according to degradation degree suffered by face image in propagation processGaussian noise kernel size in (a)Constrained toIn downsampling sizeNoise levelCompression ratio. At the same time, a small scale normal distribution is adopted to initialize the quality degradation generation algorithm with a relatively low valueIs a backbone network of (1)Ensuring that the initial output value of the backbone network approaches 0 so as to lead the quality degradation generation algorithmA non-degenerate face image is generated at an initial stage. Backbone networkMainly for predicting degradation settingsIn (a)、、、The value of (2) is as follows:
；
step three: quality degradation generation algorithm According to degradation settingsAnd the obtained original face imageGenerating an original face imageCorresponding degraded face imageObtaining a human face image pairCorresponding real tag pairs. Specifically, the quality degradation generation algorithmBased on backbone networkPredicted degradation settingsFor the original face imagePerforming degradation operation to generate a degraded face image：
，
Wherein, A convolution operation is represented and is performed,Representing scale factorsIs used for the down-sampling operation of (a),Representation ofA compressed quality factor.
Step four: based on the face image pairAdaptive depth forgery detection algorithmPerforming dual tasks of image quality evaluation and depth forgery detection to obtain corresponding original face images respectivelyAnd degrading the face imageIs a pair of outputs of (a) and (b): face image quality label pair for evaluating face image qualityAnd a detection label pair for judging the authenticity of the deeply forged face image. Specifically, the adaptive depth forgery detection algorithmIs a backbone neural network of (2)From the original face images respectivelyAnd generating an algorithm from the quality degradationGenerating corresponding degraded face imagesExtraction of human face image representationAnd. After extracting the characterization, the characterizationAndWill be respectively input to the backbone networkThe image quality evaluation branch and the depth forgery detection branch of the image quality evaluation system to obtain a face image quality label pair for evaluating the face image qualityDetection label pair for judging authenticity of deeply forged face image. In order to constrain the face image quality relationship between the original face image and the degraded face image, an adaptive depth forgery detection algorithmThe training process of the system is to introduce a face image quality sequence loss function:
，
Wherein, The number of face images is represented,AndRespectively the firstZhang Yuanshi predictive quality labels for face images and degraded face images,The superparameter defines the minimum difference between two predicted quality labels required to get zero loss. And the depth falsification detection task loss function can be expressed as:
，
Wherein, Represent the firstThe real label of the face image is opened,Represent the firstZhang Yuanshi the prediction probability of the fake face image corresponding to the face image.
Step five: label pair according to face image qualityAnd detecting tag pairsAdaptive sampling networkFace image quality feedback obtained through calculationAnd algorithm performance feedback. Specifically, according to the face image quality label, an image quality score is introducedTo represent a backbone networkThe higher the predicted image quality score, the better the image quality:
，
Wherein, Representing a gaussian distribution. At the same time, an adaptive sampling networkCalculating a real tag from a detected tagPredictive tagsAlgorithm performance feedback between. The algorithm performance feedback results provide an algorithm performance prediction error for each face image, with smaller errors representing higher confidence.
Step six: adaptive sampling networkFeedback according to human face image qualityAnd algorithm performance feedbackComputing pairs of images relating to facesTraining weights of (a)The training weightIs used to calculate a weight loss, and finally the weight loss is used to assist the quality degradation generation algorithmAnd adaptive depth forgery detection algorithmUpdating algorithm parameters. Specifically, the adaptive sampling networkFeedback according to human face image qualityAnd algorithm performance feedbackAdaptive sampling networkIs the original face imageAnd corresponding degraded face imagesAdaptive generation of corresponding weightsAnd. Weight pair based on training face imageQuality degradation generation algorithmAnd adaptive depth forgery detection algorithmThe training loss function of (2) is as follows:
，
，
Wherein, Representing the weight of the training face image,Is a non-negative super-parameter of the balance loss function.
Adaptive sampling networkAnd predicting the weight of each face image in the current training iteration by adopting a local optimal strategy, and adaptively adjusting the training weight of each face image. Adaptive sampling networkThe optimization function of (2) is as follows:
，
Wherein, Is a super parameter.
Algorithm reasoning:
Step one: obtaining a face image to be detected ；
Step two: the face image to be measuredInput to trained adaptive depth forgery detection algorithmThe self-adaptive depth forgery detection algorithmCalculating to obtain face image to be measuredIs a result of detection of deep forgery of (a)。
It should be understood that the foregoing description of the preferred embodiments is not intended to limit the scope of the invention, but rather to limit the scope of the claims, and that those skilled in the art can make substitutions or modifications without departing from the scope of the invention as set forth in the appended claims.
Claims (3)
1. The high-robustness deep forgery detection method based on self-adaptive learning is characterized by comprising the following steps of: an algorithm training phase and an algorithm reasoning phase, in particular:
Algorithm training phase:
Obtaining an original face image x and a corresponding label y;
generating a degradation setting S by a quality degradation generating algorithm G;
The quality degradation generation algorithm G generates a degradation face image x ' corresponding to the original face image x according to the degradation setting S and the obtained original face image x, and obtains a face image pair { x, x ' } and a corresponding real label pair { y, y ' };
According to the face image pair { x, x '}, the self-adaptive depth forgery detection algorithm F performs the dual tasks of image quality evaluation and depth forgery detection to obtain two output pairs corresponding to the original face image x and the degraded face image x': image quality tag pair for evaluating image quality And detecting label pair for judging image authenticityThe self-adaptive depth forging detection algorithm F extracts image representations F θ (x) and F θ (x ') from an original face image x and a corresponding degraded face image x', and a backbone network F θ which is input to the self-adaptive depth forging detection algorithm F is represented to obtain a quality label pair for evaluating image qualityAnd detecting label pair for judging image authenticityBased on quality tag pairsThe adaptive depth falsification detection algorithm F introduces a quality order loss function in the image quality assessment task:
wherein N represents the number of face images, AndThe predicted quality labels of the i Zhang Yuanshi th face image and the degraded face image are respectively, m is a super parameter, and the loss function of the depth forgery detection task can be expressed as follows:
wherein y i represents the real label of the ith face image, Representing the prediction probability of the fake face image corresponding to the i Zhang Yuanshi th face image;
tag pairs based on image quality And detecting tag pairsAdaptive sampling networkThe face image quality feedback s q and the algorithm performance feedback delta y are obtained through calculation, and the network is adaptively sampled according to the image quality labelThe image quality feedback s q is calculated to represent the predicted image quality score of the backbone network f θ, the higher the score, i.e. the better the image quality:
sq＝p(1|q)；
wherein p represents a gaussian distribution, and an adaptive sampling network Calculating true label y and predictive label from detected labelAlgorithm performance feedback betweenAdaptive sampling networkTraining weights { w, w ' } of the face image pair { x, x ' } are calculated according to algorithm performance feedback deltay and face image quality feedback s q, the training weights { w, w ' } are used for calculating weighting loss, and finally algorithm parameters are updated by utilizing the weighting loss to help a quality degradation generation algorithm G and an adaptive depth falsification detection algorithm F, and the algorithm parameters are updated based on an adaptive sampling networkThe training loss functions of the quality degradation generation algorithm G and the adaptive depth falsification detection algorithm F for the respective weights w and w 'adaptively generated for the original face image x and the respective degraded face image x' are as follows:
LG＝wxLcls，
LF＝wx(Lcls+λLorder)，
wherein w x represents training face image weights, including original face image x and corresponding degraded face image x', lambda is a non-negative hyper-parameter of the balance loss function,
Adaptive sampling networkPredicting weight of each face image in current training iteration by adopting local optimal strategy, and self-adapting sampling networkThe optimization function of (2) is as follows:
Wherein M is a super parameter;
algorithm reasoning:
Obtaining a face image x to be detected;
the method is input into a trained self-adaptive depth counterfeiting detection algorithm F, and the self-adaptive depth counterfeiting detection algorithm F calculates a classification result y of the depth counterfeiting detection.
2. The method for detecting deep forgery with high robustness based on self-adaptive learning according to claim 1, wherein the generating a degradation set S by a quality degradation generating algorithm G is specifically:
the quality degradation generation algorithm G adopts four classical degradation algorithms of Gaussian noise, gaussian blur, downsampling and JPEG compression to simulate degradation operation suffered by an image in the propagation process of a social media platform, and constrains a Gaussian blur kernel k, a downsampling size S, a Gaussian noise size alpha and a value range of a JPEG compression ratio c in degradation setting S, and a backbone network f φ of the quality degradation generation algorithm G is mainly used for predicting the value sizes of k, S, alpha and c in degradation setting S:
S＝(k,s,α,c)。
3. the method for detecting high robustness deep forgery based on adaptive learning according to claim 1, wherein the quality degradation generating algorithm G generates a degraded face image x 'corresponding to the original face image x according to the degradation setting S and the obtained original face image x, and obtains a face image pair { x, x' }, specifically:
The quality degradation generation algorithm G carries out degradation operation on the original face image x according to degradation setting S predicted by the backbone network f φ to generate a degraded face image x':
Wherein, Representing convolution operation, +.s representing downsampling operation of scale factor s, c representing JPEG compression ratio.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410834314.5A CN118379608B (en) | 2024-06-26 | 2024-06-26 | High-robustness deep forgery detection method based on self-adaptive learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410834314.5A CN118379608B (en) | 2024-06-26 | 2024-06-26 | High-robustness deep forgery detection method based on self-adaptive learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN118379608A CN118379608A (en) | 2024-07-23 |
| CN118379608B true CN118379608B (en) | 2024-10-18 |
Family
ID=91912115
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410834314.5A Active CN118379608B (en) | 2024-06-26 | 2024-06-26 | High-robustness deep forgery detection method based on self-adaptive learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN118379608B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118015715A (en) * | 2024-03-15 | 2024-05-10 | 大连海事大学 | Distributed AIGC counterfeiting detection method based on deep fusion network |
| CN118071999A (en) * | 2024-04-17 | 2024-05-24 | 厦门大学 | Multi-view 3D target detection method based on sampling self-adaption continuous NeRF |
Family Cites Families (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109829436B (en) * | 2019-02-02 | 2022-05-13 | 福州大学 | Multi-face tracking method based on depth appearance characteristics and self-adaptive aggregation network |
| CN112818767B (en) * | 2021-01-18 | 2023-07-25 | 深圳市商汤科技有限公司 | Data set generation and forgery detection methods and devices, electronic equipment and storage medium |
| CN112528969B (en) * | 2021-02-07 | 2021-06-08 | 中国人民解放军国防科技大学 | Face image authenticity detection method and system, computer equipment and storage medium |
| KR102525486B1 (en) * | 2021-04-19 | 2023-05-02 | 스냅태그 주식회사 | Method and apparatus for discriminating between original image and forgery image based on out-of-distribution detection |
| US12125256B2 (en) * | 2021-06-23 | 2024-10-22 | Intel Corporation | Generator exploitation for deepfake detection |
| CN114445891B (en) * | 2022-01-19 | 2024-08-23 | 湖南大学 | Face counterfeiting detection and positioning method based on SSIM difference graph |
| US20230401824A1 (en) * | 2022-06-08 | 2023-12-14 | Mohamed bin Zayed University of Artificial Intelligence | Video transformer for deepfake detection with incremental learning |
| CN115482595B (en) * | 2022-09-27 | 2023-04-07 | 北京邮电大学 | Specific character visual sense counterfeiting detection and identification method based on semantic segmentation |
| CN115588226A (en) * | 2022-11-01 | 2023-01-10 | 重庆邮电大学 | High-robustness deep-forged face detection method |
| CN116958846A (en) * | 2022-11-15 | 2023-10-27 | 腾讯科技（深圳）有限公司 | Video detection method, device, equipment, medium and product |
| CN116778545A (en) * | 2023-03-13 | 2023-09-19 | 中国电子科技集团公司第三十研究所 | Video depth forgery detection method and device |
| CN116110110A (en) * | 2023-03-15 | 2023-05-12 | 天津科技大学 | Fake image detection method, terminal and storage medium based on face key points |
| CN116721473A (en) * | 2023-03-23 | 2023-09-08 | 长沙理工大学 | Face counterfeiting detection method, system, equipment and medium based on frequency domain multi-scale |
| CN116206375B (en) * | 2023-04-28 | 2023-07-25 | 南京信息工程大学 | Face counterfeiting detection method based on double-layer twin network and sustainable learning |
| CN116704580A (en) * | 2023-06-09 | 2023-09-05 | 成都信息工程大学 | Face counterfeiting detection method based on depth information decoupling |
| CN117238041A (en) * | 2023-09-11 | 2023-12-15 | 西安交通大学 | Method, system, equipment and medium for detecting deep forgery of face image |
| CN117218707B (en) * | 2023-10-07 | 2024-04-16 | 南京信息工程大学 | Deep face detection method based on positive disturbance |
| CN117523439A (en) * | 2023-10-31 | 2024-02-06 | 西安理工大学 | Social network depth fake video detection method and system based on space-time characteristics |
| CN117421433B (en) * | 2023-11-17 | 2024-08-27 | 南京蓝鲸人网络科技有限公司 | Image-text intelligent public opinion analysis method and system |
| CN117975577A (en) * | 2024-01-31 | 2024-05-03 | 南京邮电大学 | Deep forgery detection method and system based on facial dynamic integration |
| CN117972610A (en) * | 2024-02-26 | 2024-05-03 | 广州佰锐网络科技有限公司 | AI-based deep pseudo detection method and system |
| CN118097502A (en) * | 2024-02-29 | 2024-05-28 | 西安理工大学 | Social network depth fake video detection method and system based on deep learning |
| CN118194342A (en) * | 2024-03-15 | 2024-06-14 | 中国科学院信息工程研究所 | Face privacy protection method for defending deep forgery |
-
2024
- 2024-06-26 CN CN202410834314.5A patent/CN118379608B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118015715A (en) * | 2024-03-15 | 2024-05-10 | 大连海事大学 | Distributed AIGC counterfeiting detection method based on deep fusion network |
| CN118071999A (en) * | 2024-04-17 | 2024-05-24 | 厦门大学 | Multi-view 3D target detection method based on sampling self-adaption continuous NeRF |
Also Published As
| Publication number | Publication date |
|---|---|
| CN118379608A (en) | 2024-07-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113554089B (en) | Image classification countermeasure sample defense method and system and data processing terminal | |
| CN110048827B (en) | Class template attack method based on deep learning convolutional neural network | |
| CN109214973B (en) | Method for generating countermeasure security carrier aiming at steganalysis neural network | |
| CN111460931A (en) | Face spoofing detection method and system based on color channel difference image characteristics | |
| CN111753881A (en) | Defense method for quantitatively identifying anti-attack based on concept sensitivity | |
| CN114492768B (en) | Twin capsule network intrusion detection method based on small sample learning | |
| CN115588226A (en) | High-robustness deep-forged face detection method | |
| CN114842267A (en) | Image classification method and system based on label noise domain self-adaption | |
| CN113538202A (en) | Image steganography method and system based on generative steganography confrontation | |
| CN112182585B (en) | Source code vulnerability detection method, system and storage medium | |
| Mazumdar et al. | Universal image manipulation detection using deep siamese convolutional neural network | |
| CN116631023A (en) | Face-changing image detection method and device based on reconstruction loss | |
| CN114330652A (en) | Target detection attack method and device | |
| CN111222583A (en) | Image steganalysis method based on confrontation training and key path extraction | |
| CN112907431B (en) | Steganalysis method for robust countersteganalysis | |
| CN118379608B (en) | High-robustness deep forgery detection method based on self-adaptive learning | |
| Rahmi et al. | Offline signature recognition using back propagation neural network | |
| CN117370832A (en) | Underwater sound target identification method and device based on Bayesian neural network | |
| CN116188439A (en) | False face-changing image detection method and device based on identity recognition probability distribution | |
| CN118587779B (en) | Living body detection method combining mask rolling and negative sample characteristic sequence | |
| CN113177599A (en) | Enhanced sample generation method based on GAN | |
| CN117854159B (en) | Face anti-fake identification method and device | |
| CN118279963B (en) | Domain adaptive face counterfeiting detection method based on meta learning architecture | |
| CN116401659B (en) | Multi-model fusion computer virus detection method based on deep learning | |
| CN114067184B (en) | Countermeasure sample detection method and system based on noise mode classification |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |