A method, system, electronic device, and storage medium for identifying risks in mobile applications.
By performing target detection and visual feature extraction on screenshots of mobile application details, and combining this with similarity judgment based on a database of disguised images, the problem of difficulty in recognizing disguised images and waste of computing resources in existing technologies has been solved, thereby improving recognition accuracy and processing speed, and extending the lifespan of recognition capabilities.
Patent Information
- Application Number
- CN202611333244.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to effectively identify high-risk mobile applications that disguise their appearance, consume significant computing resources, and struggle to balance recognition accuracy with real-time processing. Static models also struggle to cope with the problem of continuous reskinning.
By obtaining detailed screenshots of mobile applications, the system uses an object detection model to extract category labels and location information of candidate targets, crops the target area and extracts visual feature vectors, calculates the similarity distance with the disguised image vector library, and makes a dual judgment by combining risk label category and disguised image hit status. The vector library is dynamically updated to improve recognition capabilities.
It improves the accuracy and processing speed of mobile application risk identification, reduces computing resource consumption, and extends the life cycle of identification capabilities, enabling effective identification of disguised images and refined risk level determination.
Smart Images

Figure CN123116274A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of network security and mobile internet application security auditing technology, and in particular to a mobile application risk identification method, system, electronic device, and storage medium. Background Technology
[0002] In recent years, the number of mobile applications has grown rapidly with the popularization of mobile internet. As the core distribution platform connecting developers and users, app stores have become increasingly important in terms of content security review mechanisms. Among various types of non-compliant applications, high-risk applications with specific business attributes have always been the focus and challenge of the review process.
[0003] To evade platform oversight, criminals often employ sophisticated disguises for such applications. Common methods include: using calculators, weather forecasts, and other utility apps as a shell, dynamically loading specific business interfaces via cloud-controlled switches after launch; or using neutral visual elements such as the twelve zodiac animals or cartoon characters in icons and promotional materials, only intermittently displaying visual symbols with specific business attributes, such as chip-shaped graphics or roulette-style interfaces, in some detail screenshots. This highly concealed presentation makes traditional detection methods ineffective.
[0004] However, existing recognition technologies still have significant shortcomings in addressing the aforementioned challenges. Most solutions rely solely on single visual features for judgment, such as recognizing only explicit risk markers like chips and roulette wheels. For high-risk applications disguised with neutral themes like the Chinese zodiac, mythological figures, or cartoon characters, existing object detection models fail to effectively identify them due to the lack of such disguised image annotations in the training data, resulting in a persistently high false negative rate. Furthermore, existing methods typically perform global feature extraction and full vector comparison on all displayed images of the application, consuming substantial computational resources. In scenarios with tens of thousands of applications reviewed daily, it is difficult to balance recognition accuracy with real-time processing. In addition, current mainstream solutions mostly employ static model architectures, with parameters fixed after deployment, making it impossible to iterate based on new variants. This makes it difficult to cope with the adversarial evolution of such applications that continuously change their appearance and themes, leading to a gradual decline in recognition capability over time.
[0005] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention
[0006] To address the aforementioned technical problems, this invention provides a mobile application risk identification method, system, electronic device, and storage medium.
[0007] Firstly, the present invention provides a method for identifying risks in mobile applications, the technical solution of which is as follows: Obtain at least one detailed screenshot of the mobile application to be tested; Input the at least one detailed screenshot into the target detection model, and output the category label, location information and confidence score of the candidate target. The category label includes a preset risk identification category. When the confidence score exceeds the first threshold, the target region where the candidate target is located is cropped according to the location information, and the visual feature vector of the target region is extracted; Calculate the similarity distance between the visual feature vector and each vector in the camouflage image vector library, compare the similarity distance with a second threshold, and determine that the candidate target hits the camouflage image when the similarity distance is less than the second threshold; The risk level of the mobile application under test is determined based on whether there are candidate targets whose category tags belong to the risk identifier category and whether there are candidate targets that match the disguised image in at least one of the detail screenshots. Based on the risk level, the mobile application under test may be blocked or allowed to be listed on the app store.
[0008] The beneficial effects of the mobile application risk identification method of the present invention are as follows: The method of this invention obtains a screenshot of the details of the application to be tested and inputs it into a target detection model. It outputs candidate target category labels, location information and confidence scores. For candidate targets exceeding a first threshold, the target area is cropped according to the location information and visual feature vectors are extracted. The similarity distance between the vectors and each vector in the disguise image vector library is calculated and compared with a second threshold to determine the disguise image. The risk level is determined by combining the risk identifier category hit and the disguise image hit results. This method solves the problems of high false negative rate of single visual features, large computational cost of global feature extraction and full comparison, and difficulty of static models to cope with continuous skin changes. It improves recognition accuracy and processing speed, reduces computing resource consumption, and extends the life cycle of recognition capability through vector library updates.
[0009] Based on the above solution, the mobile application risk identification method of the present invention can be further improved as follows.
[0010] In one alternative approach, the step of inputting the at least one detailed screenshot into the object detection model and outputting the category label, location information, and confidence score of the candidate object includes: Each detail screenshot is input into the target detection model for forward propagation inference. The target detection model extracts features from each detail screenshot and generates a feature map. Based on the feature map, bounding box regression and classification prediction are performed on the preset anchor boxes. The category label, location information, and confidence score of each candidate target are output. The location information includes the bounding box coordinates, and the confidence score is used to characterize the probability that the candidate target belongs to the category indicated by the category label.
[0011] The advantages of adopting the above optional method are as follows: further forward propagation inference is performed on the detailed screenshot through the target detection model. After feature extraction to generate feature maps, bounding box regression and classification prediction are performed on the preset anchor boxes. The category labels, bounding box coordinates and confidence scores of the candidate targets are output, so that the location and category probability of the risk targets can be quantified, providing an accurate basis for subsequent processing.
[0012] In one alternative approach, the step of cropping the target region containing the candidate target based on the location information and extracting the visual feature vector of the target region includes: Obtain the bounding box coordinates corresponding to the candidate target from the location information, crop the target region from the at least one detail screenshot based on the bounding box coordinates, input the target region into the feature extraction network, perform convolution and pooling operations on the target region through the feature extraction network, and output the normalized visual feature vector.
[0013] The advantages of using the above-mentioned optional method are as follows: the target area is further cropped from the details screenshot based on the bounding box coordinates, and the target area is input into the feature extraction network for convolution and pooling operations to output a normalized visual feature vector. This avoids global processing of the entire image and focuses on extracting key visual information from risk areas while reducing computational overhead.
[0014] In one alternative approach, the step of calculating the similarity distance between the visual feature vector and each vector in the camouflage image vector library includes: Each of the input vectors is read sequentially from the camouflage image vector library. The cosine distance or Euclidean distance between the visual feature vector and the currently read input vector is calculated. The calculated cosine distance or Euclidean distance value is used as the similarity distance. Wherein, the visual feature vector and the input vector are both normalized vectors of the same dimension.
[0015] The advantages of adopting the above optional method are as follows: Further read the input vectors from the disguised image vector library in sequence, calculate the cosine distance or Euclidean distance between the visual feature vector and the input vector as the similarity distance, and the visual feature vector and the input vector are both normalized vectors of the same dimension, so that the distance measurement standard is unified and the possibility of misjudgment caused by vector differences is reduced.
[0016] In one optional approach, the step of determining the risk level of the mobile application under test based on whether there are candidate targets in the at least one detail screenshot whose category label belongs to the risk identifier category and whether there are candidate targets that match the disguised image includes: Iterate through all candidate targets for each detail screenshot, determine whether the category label of each candidate target belongs to the risk identifier category, and determine whether each candidate target matches the disguised image; If any screenshot of the details contains a candidate target whose category tag belongs to the risk identifier category and a candidate target that matches the disguised image, the mobile application under test is directly determined to be of a high-risk level. When each screenshot contains only candidate targets whose category tags belong to the risk identifier category and no candidate targets that match the disguised image, the risk identifier score of each candidate target is accumulated. When the accumulated result exceeds the third threshold, the mobile application under test is determined to be of high risk level. Each candidate target has a risk identifier score set according to the risk identifier category corresponding to the category tag. When each screenshot contains only candidate targets that match the disguised image and no candidate targets whose category tags belong to the risk identifier category, the image risk score of each candidate target is accumulated. When the accumulated result exceeds the third threshold, the mobile application under test is determined to be of high risk level; otherwise, the mobile application under test is determined to be of low risk level.
[0017] The beneficial effects of adopting the above optional methods are as follows: further combining the risk identification category hit and the disguise image hit for hierarchical judgment, applications that hit both features simultaneously are directly classified as high risk, and for single feature hits, a score accumulation mechanism is adopted and compared with a third threshold to achieve a refined division of risk levels, making the judgment results more flexible and accurate.
[0018] In one alternative approach, the step of performing blocking or allowing the app store to list the mobile application under test based on the risk level includes: When the mobile application under test is determined to be of high risk, it will be blocked; when the mobile application under test is determined to be of low risk, it will be allowed to be listed on the app store.
[0019] The beneficial effects of adopting the above-mentioned optional methods are as follows: further implement differentiated processing strategies based on the determined risk level, block high-risk applications to prevent illegal content from flowing into the platform, and allow low-risk applications to be released, thereby reducing obstacles to the listing of normal applications while ensuring content security, and making reasonable allocation of review resources.
[0020] In one alternative approach, it also includes: Obtain the icon image of the mobile application to be tested, compare the similarity of the icon image with each vector in the dummy image vector library, and trigger a risk warning when the similarity exceeds the fourth threshold. When the mobile application under test is determined to be of a high-risk level based on the detailed screenshot and the risk warning is triggered, the risk warning is output as additional risk evidence along with the high-risk level. When the mobile application under test is determined to be of low risk level based on the detailed screenshot but triggers the risk warning, the risk level of the mobile application under test is adjusted to medium risk level and sent for manual review.
[0021] The beneficial effects of adopting the above optional methods are as follows: further introducing similarity comparison between icon images and camouflaged image vector libraries as an auxiliary detection method, superimposing icon risk warnings on the basis of detailed screenshot judgment, providing additional evidence support for high-risk judgment, upgrading low-risk but icon abnormal situations to medium-risk and pushing them for manual review, and reducing the probability of missed detection by a single detection method.
[0022] Secondly, this invention provides a mobile application risk identification system, the technical solution of which is as follows: The acquisition module is used to acquire at least one detailed screenshot of the mobile application under test; The detection module is used to input the at least one detailed screenshot into the target detection model and output the category label, location information and confidence score of the candidate target, wherein the category label includes a preset risk identification category; The extraction module is used to crop the target region where the candidate target is located based on the location information and extract the visual feature vector of the target region when the confidence score exceeds a first threshold. The comparison module is used to calculate the similarity distance between the visual feature vector and each vector in the camouflage image vector library, compare the similarity distance with a second threshold, and determine that the candidate target hits the camouflage image when the similarity distance is less than the second threshold. The determination module is used to determine the risk level of the mobile application under test based on whether there are candidate targets whose category tags belong to the risk identifier category and whether there are candidate targets that match the disguised image in the at least one detail screenshot; The execution module is used to perform interception processing or allow listing processing on the mobile application under test according to the risk level.
[0023] The beneficial effects of the mobile application risk identification system of the present invention are as follows: The system of this invention acquires a screenshot of the details of the application to be tested and inputs it into a target detection model. It outputs candidate target category labels, location information, and confidence scores. For candidate targets exceeding a first threshold, it crops the target region based on the location information and extracts visual feature vectors. It calculates the similarity distance between the vectors and the vectors in the disguise image vector library and compares them with a second threshold to determine the disguise image. It determines the risk level by combining the dual judgment results of the risk identifier category hit and the disguise image hit. This system solves the problems of high false negative rate of single visual features, large computational cost of global feature extraction and full comparison, and difficulty of static models in dealing with continuous skin changes. It improves recognition accuracy and processing speed, reduces computational resource consumption, and extends the life cycle of recognition capability through vector library updates.
[0024] Thirdly, the technical solution of an electronic device according to the present invention is as follows: The device includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the mobile application risk identification method of the present invention.
[0025] Fourthly, the technical solution of a computer-readable storage medium provided by the present invention is as follows: The computer-readable storage medium stores instructions that, when read, cause the computer-readable storage medium to perform the steps of the mobile application risk identification method of the present invention.
[0026] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other objects, features and advantages of the present invention more obvious and understandable, specific embodiments of the present invention are described below. Attached Figure Description
[0027] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating an embodiment of a mobile application risk identification method according to the present invention. Figure 2 This is a schematic diagram of the cascaded comparison process; Figure 3 This is a schematic diagram of the comprehensive risk assessment logic; Figure 4 A schematic diagram of the self-evolutionary feedback update process; Figure 5 This is a schematic diagram of the structure of an embodiment of the mobile application risk identification system of the present invention; Figure 6This is a schematic diagram of an embodiment of an electronic device according to the present invention. Detailed Implementation
[0028] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.
[0029] Figure 1 This diagram illustrates a flowchart of an embodiment of a mobile application risk identification method provided by the present invention. This method can be executed by an electronic device such as a terminal device or a server. The terminal device can be any fixed or mobile terminal, such as a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, in-vehicle device, or wearable device. The server can be a single server or a server cluster consisting of multiple servers. Any electronic device can implement the mobile application risk identification method by having its processor call computer-readable instructions stored in its memory. Figure 1 As shown, it includes the following steps: S1. Obtain at least one detailed screenshot of the mobile application to be tested.
[0030] The term "mobile application under test" refers to an application that needs to be reviewed and approved by an app store before it can be listed for download and installation by users. For example, app store reviewers might designate a mobile application named "Mobile Application A" as the application under test and determine whether it should be allowed to be listed through risk identification. The "detail screenshot" refers to an image displayed on the app store's showcase page that presents the application's interface to users. For example, Mobile Application A might provide a detail screenshot on its showcase page, which simultaneously displays an icon of a dragon (one of the twelve Chinese zodiac animals) and a roulette game interface labeled "Lucky Wheel."
[0031] S2. Input the at least one detailed screenshot into the target detection model and output the category label, location information and confidence score of the candidate target. The category label includes a preset risk identification category.
[0032] The object detection model refers to a deep learning network model that has been pre-trained to detect target objects in an input image and output the target category and location. For example, after inputting a screenshot of the details of mobile application A into the object detection model, the object detection model outputs the detection result for the wheel area in the screenshot.
[0033] Here, a candidate target refers to a target region that the target detection model detects from the input image that may contain a specific object; for example, the target detection model detects a candidate target from a screenshot of the details of mobile application A, and the candidate target corresponds to the wheel interface region in the screenshot of the details.
[0034] Here, category label refers to the tag information used to identify the category to which the candidate target belongs; for example, the category label output by the target detection model for the candidate target detected in the details screenshot of mobile application A is "wheel". Location information refers to the spatial coordinate data used to describe the location of the candidate target in the details screenshot; for example, the location information output by the target detection model for the wheel candidate target detected in the details screenshot of mobile application A is the bounding box coordinates. The bounding box coordinates indicate that the wheel is located at the top left corner of the details screenshot. The coordinates of the bottom right corner are Within the rectangular area. The confidence score refers to the probability value output by the object detection model that the candidate object belongs to the category indicated by the category label; for example, the confidence score output by the object detection model for the roulette candidate object detected in the details screenshot of mobile application A is 0.92, which means that the confidence level of the roulette detection result is 92%.
[0035] Among them, the risk label category refers to the category of visual elements that are pre-defined and associated with the risky behavior of mobile applications; for example, "roulette", "chips", "playing cards" and "dice" are preset as risk label categories, and the category label "roulette" of the roulette candidate target detected in the details screenshot of mobile application A belongs to the risk label category.
[0036] S3. When the confidence score exceeds the first threshold, the target region where the candidate target is located is cropped according to the location information, and the visual feature vector of the target region is extracted.
[0037] The first threshold refers to the minimum confidence score threshold used to filter the output results of the target detection model. For example, the first threshold is set to 0.45. In the screenshot of the details of mobile application A, the confidence score of the candidate target in the roulette wheel is 0.92, which is greater than the first threshold of 0.45. The candidate target in the roulette wheel passes the screening and enters the subsequent processing flow.
[0038] Here, the target region refers to: an image sub-region cropped from the details screenshot based on the location information of the candidate targets; for example, based on the location information of the wheel candidate targets in the details screenshot of mobile application A. The rectangular image region containing the wheel is cropped from the details screenshot as the target region. A visual feature vector refers to the numerical vector representation obtained after transforming the image content of the target region through a feature extraction network. For example, after inputting the target region of the wheel cropped from the details screenshot of mobile application A into the feature extraction network, the network outputs a 512-dimensional normalized visual feature vector.
[0039] S4. Calculate the similarity distance between the visual feature vector and each vector in the camouflage image vector library, compare the similarity distance with a second threshold, and determine that the candidate target hits the camouflage image when the similarity distance is less than the second threshold.
[0040] The "disguised image vector library" refers to a pre-built database containing standard feature vectors for multiple disguised image categories, with each category corresponding to a standard input vector. For example, the library contains the standard feature vector for the dragon image from the Chinese zodiac. The visual feature vector extracted from the dragon image next to the target area on the wheel in the screenshot of mobile application A will be compared with the standard vector for the dragon image. The "input vector" refers to the standard feature vector corresponding to each disguised image category stored in the library. For example, the library stores the standard input vector for the dragon image from the Chinese zodiac, and this standard input vector is used to calculate similarity with the visual feature vectors of subsequently detected candidate targets.
[0041] The similarity distance refers to a numerical value used to measure the degree of feature difference between the visual feature vector and the input vector in the database. For example, the cosine distance between the visual feature vector of the dragon image target area in the screenshot of mobile application A's details and the input vector of the dragon image in the disguise image vector library is calculated, resulting in a similarity distance of 0.15. The second threshold refers to the maximum allowable similarity distance used to determine whether a candidate target matches a disguised image. For example, the second threshold is set to 0.25. Since the similarity distance of 0.15 between the dragon image target area in the screenshot of mobile application A's details and the input vector of the dragon image in the database is less than the second threshold of 0.25, the candidate target is determined to match a disguised image.
[0042] Among them, the disguised image refers to the neutral theme visual elements adopted by mobile applications to avoid risk review. Neutral theme visual elements include the images of the twelve zodiac animals, mythological figures, and cartoon characters. For example, the dragon image of the twelve zodiac animals that appears in the details screenshot of mobile application A is a disguised image. The dragon image itself does not have risk identification attributes, but it is often used by high-risk applications for packaging and disguise.
[0043] S5. Determine the risk level of the mobile application under test based on whether there are candidate targets whose category labels belong to the risk identifier category and whether there are candidate targets that match the disguised image in the at least one detail screenshot.
[0044] The risk level refers to the risk classification result of the mobile application under test according to the risk assessment rules. For example, if the screenshot of the details of mobile application A shows both the risk label category of a roulette wheel and the disguised image of the Chinese zodiac dragon, mobile application A is judged to be of high risk level.
[0045] S6. Based on the risk level, perform interception or allow listing processing on the mobile application under test.
[0046] Interception refers to the control operation of prohibiting the mobile application under test from being listed on the app store. For example, if mobile application A is judged to be of a high-risk level, the app store will intercept mobile application A, and mobile application A will not pass the review and be listed. Allowing listing refers to the control operation of allowing the mobile application under test to be listed and published in the app store. For example, if another mobile application named mobile application B does not detect any risk label category or match any disguise image after risk identification, mobile application B is judged to be of a low-risk level, and the app store will allow mobile application B to be listed.
[0047] The technical solution of this embodiment obtains a screenshot of the details of the application to be tested and inputs it into the target detection model. It outputs candidate target category labels, location information and confidence scores. For candidate targets that exceed the first threshold, the target area is cropped according to the location information and visual feature vectors are extracted. The similarity distance between the vectors and each vector in the disguise image vector library is calculated and compared with the second threshold to determine the disguise image. The risk level is determined by combining the risk identifier category hit and the disguise image hit results. This solves the problems of high false negative rate of single visual features, large computational cost of global feature extraction and full comparison, and difficulty of static models to cope with continuous skin changes. It improves recognition accuracy and processing speed, reduces computing resource consumption, and extends the life cycle of recognition capability through vector library updates.
[0048] In one alternative approach, the step of inputting the at least one detailed screenshot into the object detection model and outputting the category label, location information, and confidence score of the candidate object includes: Each detail screenshot is input into the target detection model for forward propagation inference. The target detection model extracts features from each detail screenshot and generates a feature map. Based on the feature map, bounding box regression and classification prediction are performed on the preset anchor boxes. The category label, location information, and confidence score of each candidate target are output. The location information includes the bounding box coordinates, and the confidence score is used to characterize the probability that the candidate target belongs to the category indicated by the category label.
[0049] Forward propagation inference refers to the computational process of passing input data from the input layer of the object detection model to the output layer layer by layer. For example, after inputting a screenshot of the details of mobile application A into the object detection model, the data propagates forward from the input layer through each convolutional layer and fully connected layer, finally obtaining the detection result at the output layer. Feature maps refer to two-dimensional matrix data generated by the object detection model during feature extraction to represent semantic information at different levels of the image. For example, after extracting features from the screenshot of the details of mobile application A, the object detection model generates feature maps at multiple scales. These feature maps are then used for subsequent bounding box regression and classification prediction of the candidate objects in the wheel.
[0050] Among them, the preset anchor boxes refer to candidate boxes with different sizes and aspect ratios predefined by the object detection model. The candidate boxes are used to slide and scan on the feature map to match targets of different sizes. For example, the object detection model lays out multiple preset anchor boxes on the feature map of the details screenshot of mobile application A. One of the preset anchor boxes with a size of 128 pixels by 128 pixels is the best match for the actual bounding box of the candidate target in the wheel.
[0051] Bounding box regression refers to the calculation process of adjusting the position of a preset anchor box to more accurately match the actual bounding box position of the candidate target. For example, the target detection model performs bounding box regression on the preset anchor boxes matching the wheel candidate targets in the details screenshot of mobile application A, adjusting the position of the preset anchor boxes from... Adjust to This allows the bounding box to more accurately encompass the wheel area. Classification prediction refers to the calculation process of determining the category to which a candidate target belongs; for example, the target detection model performs classification prediction on the candidate targets in the details screenshot of mobile application A after bounding box regression adjustment, outputting the category label of the candidate target as "wheel" and assigning a confidence score of 0.92.
[0052] Wherein, the bounding box coordinates refer to the four-dimensional coordinate data used to uniquely determine the location of the rectangular area where the candidate target is located in the details screenshot; for example, in the details screenshot of mobile application A, the bounding box coordinates of the candidate targets output by the wheel are: The bounding box coordinates precisely define the position of the wheel in the details screenshot.
[0053] In the above-mentioned optional methods, forward propagation inference is further performed on the detailed screenshot through the target detection model. After feature extraction to generate a feature map, bounding box regression and classification prediction are performed on the preset anchor box. The category label, bounding box coordinates and confidence score of the candidate target are output, so that the location and category probability of the risk target can be quantified, providing an accurate basis for subsequent processing.
[0054] In one alternative approach, the step of cropping the target region containing the candidate target based on the location information and extracting the visual feature vector of the target region includes: Obtain the bounding box coordinates corresponding to the candidate target from the location information, crop the target region from the at least one detail screenshot based on the bounding box coordinates, input the target region into the feature extraction network, perform convolution and pooling operations on the target region through the feature extraction network, and output the normalized visual feature vector.
[0055] The feature extraction network refers to a deep learning network model used to convert an input image into a numerical feature vector representation. For example, the target region of the Chinese zodiac dragon cropped from the details screenshot of mobile application A is input into the feature extraction network. After multi-layer convolution operations and pooling processing, the feature extraction network outputs a 512-dimensional normalized visual feature vector.
[0056] In the above-mentioned optional methods, the target area is further cropped from the details screenshot based on the bounding box coordinates. The target area is then input into the feature extraction network for convolution and pooling operations, and a normalized visual feature vector is output. This avoids global processing of the entire image, reducing computational overhead while focusing on extracting key visual information from risky areas.
[0057] In one alternative approach, the step of calculating the similarity distance between the visual feature vector and each vector in the camouflage image vector library includes: Each of the input vectors is read sequentially from the camouflage image vector library. The cosine distance or Euclidean distance between the visual feature vector and the currently read input vector is calculated. The calculated cosine distance or Euclidean distance value is used as the similarity distance. Wherein, the visual feature vector and the input vector are both normalized vectors of the same dimension.
[0058] Cosine distance measures the similarity between two vectors by calculating the cosine of the angle between them. For example, calculating the cosine distance between the visual feature vector of the dragon target area in the screenshot of mobile application A and the dragon image vector in the disguised image vector library yields a cosine distance of 0.15. A smaller cosine distance indicates greater similarity between the two vectors. Euclidean distance measures the straight-line distance between two vectors in Euclidean space. For example, calculating the Euclidean distance between the visual feature vector of the dragon target area in the screenshot of mobile application A and the dragon image vector in the disguised image vector library yields a Euclidean distance of 0.32. A smaller Euclidean distance indicates greater proximity between the two vectors.
[0059] In the above optional methods, the input vectors are read sequentially from the camouflage image vector library, and the cosine distance or Euclidean distance between the visual feature vector and the input vector is calculated as the similarity distance. Furthermore, the visual feature vector and the input vector are both normalized vectors of the same dimension, so that the distance measurement standard is unified and the possibility of misjudgment caused by vector differences is reduced.
[0060] In one optional approach, the step of determining the risk level of the mobile application under test based on whether there are candidate targets in the at least one detail screenshot whose category label belongs to the risk identifier category and whether there are candidate targets that match the disguised image includes: Iterate through all candidate targets for each detail screenshot, determine whether the category label of each candidate target belongs to the risk identifier category, and determine whether each candidate target matches the disguised image.
[0061] If any screenshot of the details contains a candidate target whose category label belongs to the risk identifier category and a candidate target that matches the disguised image, the mobile application under test is directly determined to be of a high-risk level.
[0062] The "high-risk level" refers to the degree of risk classification of the mobile application under test, which indicates that the application has obvious risky behavior and should be prohibited from being listed. For example, mobile application A is classified as high-risk because it shows both the risky category of a roulette wheel and the disguised image of the Chinese zodiac dragon in its details screenshot. The app store will then block mobile application A.
[0063] When each screenshot contains only candidate targets whose category tags belong to the risk identifier category and no candidate targets that match the disguised image, the risk identifier score of each candidate target is accumulated. When the accumulated result exceeds the third threshold, the mobile application under test is determined to be of high risk level. Each candidate target has a risk identifier score set according to the risk identifier category corresponding to the category tag.
[0064] The risk label score refers to a weighted score based on the risk label category of the candidate target. Different risk label categories correspond to different risk label scores. For example, the risk label score for a roulette wheel is 15 points, for chips it is 10 points, for playing cards it is 8 points, and for dice it is 6 points. The third threshold is a scoring threshold used to determine whether detecting only a risk label category or only hitting a disguised image constitutes a high-risk level. For example, the third threshold is set to 25 points. If a mobile application detects only two candidate targets, roulette wheel and chips, the accumulated risk label score is 15 points plus 10 points equals 25 points. The accumulated result equals the third threshold of 25 points, and is judged as a high-risk level.
[0065] When each screenshot contains only candidate targets that match the disguised image and no candidate targets whose category tags belong to the risk identifier category, the image risk score of each candidate target is accumulated. When the accumulated result exceeds the third threshold, the mobile application under test is determined to be of high risk level; otherwise, the mobile application under test is determined to be of low risk level.
[0066] The image risk score refers to the weighted score assigned when a candidate target is determined to be a disguised image. It is used to accumulate the image risk score in the case where no risk identifier category is detected. Each candidate target that is a disguised image is assigned a fixed image risk score. For example, in the details screenshot of mobile application A, a disguised image of a dragon from the Chinese zodiac and a disguised image of a cartoon character are found. The image risk score of each disguised image is 20 points. The image risk score of the two disguised images is accumulated to 40 points. After comparing with the third threshold of 25 points, it exceeds the third threshold. Therefore, mobile application A is judged to be of high risk level.
[0067] The low-risk level refers to the risk level classification result of the mobile application under test not having any obvious risky behavior and should be allowed to be listed. For example, after risk identification, the screenshot of the details of a mobile application named Mobile Application B does not detect any risk label category or match any disguised image. Mobile Application B is judged to be low-risk, and the app store will allow Mobile Application B to be listed.
[0068] Among the above-mentioned optional methods, a further classification judgment is made by combining the hit status of risk identification category and the hit status of disguise image. Applications that hit both features at the same time are directly classified as high risk. For cases where a single feature is hit, a score accumulation mechanism is adopted and compared with a third threshold to achieve a refined classification of risk level, making the judgment results more flexible and accurate.
[0069] In one alternative approach, the step of performing blocking or allowing the app store to list the mobile application under test based on the risk level includes: When the mobile application under test is determined to be of high risk, it will be blocked; when the mobile application under test is determined to be of low risk, it will be allowed to be listed on the app store.
[0070] Among the above-mentioned optional methods, a differentiated processing strategy is further implemented based on the determined risk level. High-risk applications are blocked to prevent illegal content from flowing into the platform, while low-risk applications are allowed to enter. This ensures content security while reducing obstacles to the listing of normal applications and allows for the rational allocation of review resources.
[0071] In one alternative approach, it also includes: Obtain the icon image of the mobile application to be tested, compare the similarity of the icon image with each vector in the dummy image vector library, and trigger a risk warning when the similarity exceeds the fourth threshold.
[0072] Here, the icon image refers to the application icon image displayed by the mobile application under test in the app store; for example, in addition to providing a screenshot of the details, mobile application A also provides an icon image containing the image of the dragon in the Chinese zodiac on the app store display page. The icon image is obtained separately for spoofing comparison.
[0073] The fourth threshold refers to the similarity comparison threshold used to determine whether an icon image triggers a risk warning. For example, the fourth threshold is set to 0.80. If the similarity between the icon image of mobile application A and the dragon image vector in the disguised image vector library is 0.85, and this similarity exceeds the fourth threshold of 0.80, a risk warning is triggered. A risk warning is a warning signal triggered when the similarity between the icon image and any vector in the disguised image vector library exceeds the fourth threshold. For example, if the similarity between the icon image of mobile application A and the dragon image vector in the disguised image vector library is 0.85, and this similarity exceeds the fourth threshold of 0.80, a risk warning is triggered.
[0074] When the mobile application under test is determined to be of a high-risk level based on the detailed screenshot and the risk warning is triggered, the risk warning is output as additional risk evidence along with the high-risk level.
[0075] When the mobile application under test is determined to be of low risk level based on the detailed screenshot but triggers the risk warning, the risk level of the mobile application under test is adjusted to medium risk level and sent for manual review.
[0076] The medium-risk level refers to the risk level classification result where the mobile application under test shows suspicious signs but is not enough to be directly judged as high-risk. The medium-risk level needs to be further confirmed by manual review. For example, the screenshot of the details of mobile application A does not detect any risk identification category, but the similarity between the icon image of mobile application A and the dragon image vector in the disguise image vector library exceeds the fourth threshold and triggers a risk warning. The risk level of mobile application A is adjusted to medium-risk level and pushed to manual review.
[0077] Among the above optional methods, a similarity comparison between icon images and a camouflaged image vector library is further introduced as an auxiliary detection method. Based on the judgment of the details screenshot, an icon risk warning is superimposed to provide additional evidence support for the high-risk judgment. For low-risk cases but with abnormal icons, the case is upgraded to medium risk and pushed for manual review, thereby reducing the probability of missed detection by a single detection method.
[0078] In this embodiment, it should be noted that for the first to fourth thresholds: The first threshold is determined based on the trade-off between the confidence score distribution and detection precision of the target detection model on the validation set. By statistically analyzing the confidence distribution of positive samples and the false detection rate of negative samples in the validation set, a confidence threshold that balances the detection recall and precision is selected as the first threshold. The value of the first threshold ranges from 0.3 to 0.6. For example, in a specific embodiment, the default value of the first threshold is 0.45.
[0079] The second threshold is determined based on the statistical distribution of similarity distances between each input vector in the camouflaged image vector library and a large number of positive and negative sample visual feature vectors. By calculating the similarity distance distribution between positive samples (known camouflaged images) and their corresponding input vectors, and the similarity distance distribution between negative samples (non-camouflaged images) and all input vectors, the distance threshold that maximizes the distinction between positive and negative samples is selected as the second threshold. The value range of the second threshold is 0.7 to 0.95. For example, in a specific embodiment, the default value of the second threshold is 0.85.
[0080] The third threshold is determined based on the correspondence between the accumulated risk label scores and the manual review judgment results. It is achieved by accumulating scores for samples that only detect risk label categories and samples that only match the disguised image in historical review data, and statistically analyzing their high-risk judgment ratio. The score threshold that makes both the false positive rate and the false negative rate within an acceptable range is selected as the third threshold. The value range of the third threshold is 15 to 40 points. For example, in a specific embodiment, the default value of the third threshold is 25 points.
[0081] The fourth threshold is determined based on the statistical distribution of similarity between the icon image and each vector in the spoofing image vector library, as well as the business's requirements for icon risk sensitivity. By performing similarity comparison statistical analysis on a large number of known spoofing application icon images, the similarity threshold that can effectively recall spoofing icon samples while ensuring a low false positive rate is selected as the fourth threshold. The value range of the fourth threshold is 0.7 to 0.9. For example, in a specific embodiment, the default value of the fourth threshold is 0.8.
[0082] like Figure 2 As shown, after the cascaded comparison process begins, a detailed screenshot and icon image of the mobile application under test are acquired, and the detailed screenshot is input into the object detection model. The object detection model incorporates a deep learning object detection model trained with transfer fine-tuning on a high-risk scene annotation dataset. This model uses a YOLO series network or an SSD network as its basic architecture. The high-risk scene annotation dataset contains over 100,000 annotated images, covering bounding boxes and category labels for chips, roulette wheels, playing cards, and dice under different resolutions, lighting conditions, and occlusion levels. The object detection model infers from the detailed screenshot and outputs the category labels and bounding box coordinates of all detected candidate objects. The bounding box coordinates precisely define the position of the candidate target in the details screenshot, and the confidence score is a floating-point number between 0 and 1. The confidence score of each candidate target is compared with a preset first threshold (range 0.3 to 0.6, default 0.45). Candidate targets with confidence scores below the first threshold are considered low-confidence detection results and are not processed; only the detected risk label category information is recorded, the process ends, and no risk is determined. For candidate targets with confidence scores reaching or exceeding the first threshold, the bounding box region of the candidate target is cropped as the target region and sent for feature extraction and vector comparison. A ResNet or EfficientNet network is used as the feature extraction network to convert the target region into a 512-dimensional normalized visual feature vector. Then, each input vector is sequentially read from the camouflage image vector library, and the cosine distance or Euclidean distance between the visual feature vector and the currently read input vector is calculated as the similarity distance.
[0083] It should be noted that the construction method of the disguised image vector library is as follows: A manual review team identifies disguised image categories from a large number of high-risk application samples (such as gambling application samples), including but not limited to the twelve zodiac animals (rat, ox, tiger, rabbit, dragon, snake, horse, sheep, monkey, rooster, dog, pig), ancient Greek mythological figures (Zeus, Hera, Poseidon, etc.), cartoon characters (Galo, clown, etc.), and specific symbol categories. For each category, at least 100 high-quality seed images are collected, and a 512-dimensional normalized feature vector is generated for each seed image. The average of all feature vectors for the same category is used as the standard input vector for that category in the disguised image vector library. All input vectors are stored in the form of a category-vector mapping table. The similarity distance is compared with a preset second threshold (range 0.7 to 0.95, default 0.85). When the similarity distance is less than the second threshold, the candidate target is determined to have matched a disguised image. When the similarity distance is greater than or equal to the second threshold, the candidate target is determined not to have matched a disguised image; only the detected risk identifier category information is recorded, the process ends, and no risk is determined. A comprehensive risk score is performed on candidate targets that have been identified in a disguised form, along with the results of the risk identification category.
[0084] like Figure 3As shown, the comprehensive risk assessment logic is as follows: It receives the risk identifier category detection results, the disguised image hit results, and the similarity comparison results between the icon image and each vector in the disguised image vector library. If at least one candidate target belonging to the risk identifier category (in this embodiment, gambling elements are used as an example by default, but adjustments can be made according to actual circumstances; no restrictions are set here) is detected in any detail screenshot and at least one disguised image is hit, the mobile application under test is directly determined to be of a high-risk level. If only the risk identifier category is detected but no disguised image is hit, the risk identifier scores of each detected candidate target (10 points for chips, 15 points for roulette, 8 points for playing cards, and 6 points for dice) are accumulated to obtain a basic risk score, which is compared with the third threshold (range 15 to 40 points, default 25 points). If the score exceeds the threshold, it is determined to be of a high-risk level. If only a disguised image is hit but no risk identifier category is detected, the image risk score of each hit disguised image (20 points for each disguised image) is accumulated to obtain an image risk score, which is compared with the third threshold. If the score exceeds the threshold, it is determined to be of a high-risk level. Icon images are used as a separate criterion. The similarity of the icon image with each vector in the masquerade image vector library is compared. If the similarity exceeds the fourth threshold (range 0.7 to 0.9, default 0.8), a risk warning is triggered. When a high-risk level is determined based on the screenshot details and a risk warning is triggered, the risk warning is output as additional risk evidence along with the high-risk level. When a low-risk level is determined based on the screenshot details but a risk warning is triggered, the risk level is adjusted to medium risk and sent for manual review. Mobile applications judged as high-risk are blocked. Those with a basic risk score or image risk score not exceeding the third threshold are initially judged as low-risk. Those that do not detect any risk marker categories, do not match any masquerade images, and do not trigger icon risk warnings are judged as safe and allowed to be listed.
[0085] like Figure 4As shown, the self-evolutionary feedback update process is as follows: In the missed detection sample mining stage, suspected missed detection samples are obtained in parallel through four channels: receiving missed detection cases submitted by app store operators; collecting app icons and detail screenshots from high-risk app promotion platforms and websites via web crawlers; querying other apps under the same developer based on the developer's identity identifier of confirmed high-risk apps; and periodically lowering the first threshold to 0.30 to broadly scan historically safe apps to obtain boundary samples. In the manual annotation and review stage, the obtained suspected samples are summarized and pushed to the manual review team for confirmation and annotation. In the knowledge base and model update stage, based on the review results, if a new disguised image category is discovered, a feature vector is constructed for that category and added to the disguised image vector library; if it is a variant of an existing disguised image, the new sample is added to the corresponding category's sample dataset. The backbone network parameters of the target detection model are frozen, and only the detection head part is fine-tuned using the new samples with a small number of iterations to enhance the target detection model's ability to identify variants of disguised images. During the dynamic threshold adjustment phase, the historical number of interceptions, false positive rate, and false negative rate are recorded for each business scenario. Optimal threshold combinations are calculated periodically for different app store platforms, review periods, and geographic sources. The values of the first, second, third, and fourth thresholds are dynamically adjusted. Threshold configurations for each scenario are independently stored in the configuration center and support online hot updates. Through the cyclical execution of these phases, the identification system can continuously evolve, developing a rapid response capability to emerging spoofing techniques.
[0086] The technical solution in this embodiment is applicable to large-scale automated review scenarios in app stores. Taking a typical scenario of tens of thousands of newly submitted and updated apps daily, the system uses an application programming interface (API) to batch acquire icon images and screenshots of apps awaiting review. After automated processing as described above, it outputs the risk level and evidence of risk detection for each app. This evidence includes the category and location of the detected risk markers, the category of the detected disguised appearance, and the similarity distance. High-risk apps are automatically blocked from listing and their review reports are pushed to the manual review queue. Low-risk apps are listed normally, and medium-risk apps are pushed to the manual review queue. Daily operational reports are generated for the review team to evaluate the identification effectiveness and adjust strategies.
[0087] Figure 5 A schematic diagram of an embodiment of a mobile application risk identification system 200 provided by the present invention is shown. Figure 5 As shown, the mobile application risk identification system 200 includes: The acquisition module 201 is used to acquire at least one detailed screenshot of the mobile application under test; Detection module 202 is used to input the at least one detailed screenshot into the target detection model and output the category label, location information and confidence score of the candidate target, wherein the category label includes a preset risk identification category; The extraction module 203 is used to crop the target region where the candidate target is located based on the location information and extract the visual feature vector of the target region when the confidence score exceeds the first threshold. The comparison module 204 is used to calculate the similarity distance between the visual feature vector and each vector in the camouflage image vector library, compare the similarity distance with a second threshold, and determine that the candidate target hits the camouflage image when the similarity distance is less than the second threshold. The determination module 205 is used to determine the risk level of the mobile application under test based on whether there are candidate targets whose category tags belong to the risk identifier category and whether there are candidate targets that match the disguised image in the at least one detail screenshot; The execution module 206 is used to perform interception processing or allow listing processing on the mobile application under test according to the risk level.
[0088] In one alternative embodiment, the detection module 202 is specifically used for: Each detail screenshot is input into the target detection model for forward propagation inference. The target detection model extracts features from each detail screenshot and generates a feature map. Based on the feature map, bounding box regression and classification prediction are performed on the preset anchor boxes. The category label, location information, and confidence score of each candidate target are output. The location information includes the bounding box coordinates, and the confidence score is used to characterize the probability that the candidate target belongs to the category indicated by the category label.
[0089] In an alternative embodiment, the extraction module 203 is specifically used for: Obtain the bounding box coordinates corresponding to the candidate target from the location information, crop the target region from the at least one detail screenshot based on the bounding box coordinates, input the target region into the feature extraction network, perform convolution and pooling operations on the target region through the feature extraction network, and output the normalized visual feature vector.
[0090] In an alternative embodiment, the comparison module 204 is specifically used for: Each of the input vectors is read sequentially from the camouflage image vector library. The cosine distance or Euclidean distance between the visual feature vector and the currently read input vector is calculated. The calculated cosine distance or Euclidean distance value is used as the similarity distance. Wherein, the visual feature vector and the input vector are both normalized vectors of the same dimension.
[0091] In one alternative embodiment, the determination module 205 is specifically used for: Iterate through all candidate targets for each detail screenshot, determine whether the category label of each candidate target belongs to the risk identifier category, and determine whether each candidate target matches the disguised image; If any screenshot of the details contains a candidate target whose category tag belongs to the risk identifier category and a candidate target that matches the disguised image, the mobile application under test is directly determined to be of a high-risk level. When each screenshot contains only candidate targets whose category tags belong to the risk identifier category and no candidate targets that match the disguised image, the risk identifier score of each candidate target is accumulated. When the accumulated result exceeds the third threshold, the mobile application under test is determined to be of high risk level. Each candidate target has a risk identifier score set according to the risk identifier category corresponding to the category tag. When each screenshot contains only candidate targets that match the disguised image and no candidate targets whose category tags belong to the risk identifier category, the image risk score of each candidate target is accumulated. When the accumulated result exceeds the third threshold, the mobile application under test is determined to be of high risk level; otherwise, the mobile application under test is determined to be of low risk level.
[0092] In an alternative embodiment, the execution module 206 is specifically used for: When the mobile application under test is determined to be of high risk, it will be blocked; when the mobile application under test is determined to be of low risk, it will be allowed to be listed on the app store.
[0093] In an alternative embodiment, the execution module 206 is further configured to: Obtain the icon image of the mobile application to be tested, compare the similarity of the icon image with each vector in the dummy image vector library, and trigger a risk warning when the similarity exceeds the fourth threshold. When the mobile application under test is determined to be of a high-risk level based on the detailed screenshot and the risk warning is triggered, the risk warning is output as additional risk evidence along with the high-risk level. When the mobile application under test is determined to be of low risk level based on the detailed screenshot but triggers the risk warning, the risk level of the mobile application under test is adjusted to medium risk level and sent for manual review.
[0094] It should be noted that the beneficial effects of the mobile application risk identification system 200 provided in the above embodiments are the same as those of the mobile application risk identification method described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.
[0095] The mobile application risk identification system 200 of the present invention can be a computer program (including program code) running on a computer device. For example, the mobile application risk identification system 200 of the present invention is an application software that can be used to execute the corresponding steps in the mobile application risk identification method of the present invention.
[0096] In some embodiments, the mobile application risk identification system 200 of the present invention can be implemented in a combination of hardware and software. As an example, the mobile application risk identification system 200 of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the mobile application risk identification method of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.
[0097] The modules described in the embodiments of this invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.
[0098] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the mobile application risk identification methods described above. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the mobile application risk identification method shown in any embodiment of the present invention by calling the computer program.
[0099] In one alternative embodiment, an electronic device is provided, such as Figure 6 As shown, Figure 6 The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.
[0100] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.
[0101] Bus 4002 may include a path for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus 4002 is represented by only one thick line, but this does not mean that there is only one bus or one type of bus.
[0102] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.
[0103] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.
[0104] Among them, electronic devices can also be terminal devices. A terminal device can be any terminal device that can install applications and access web pages through applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.
[0105] It should be noted that, Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.
[0106] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the mobile application risk identification methods described above.
[0107] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.
[0108] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned mobile application risk identification method.
[0109] Computer program code for performing the operations of this invention can be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, and C++, and conventional procedural programming languages such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0110] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0111] The computer-readable storage medium provided in this invention can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
[0112] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.
[0113] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.
[0114] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.
[0115] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.
[0116] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A method for identifying risks in mobile applications, characterized in that, include: Obtain at least one detailed screenshot of the mobile application to be tested; Input the at least one detailed screenshot into the target detection model, and output the category label, location information and confidence score of the candidate target. The category label includes a preset risk identification category. When the confidence score exceeds the first threshold, the target region where the candidate target is located is cropped according to the location information, and the visual feature vector of the target region is extracted; Calculate the similarity distance between the visual feature vector and each vector in the camouflage image vector library, compare the similarity distance with a second threshold, and determine that the candidate target hits the camouflage image when the similarity distance is less than the second threshold; The risk level of the mobile application under test is determined based on whether there are candidate targets whose category tags belong to the risk identifier category and whether there are candidate targets that match the disguised image in at least one of the detail screenshots. Based on the risk level, the mobile application under test may be blocked or allowed to be listed on the app store.
2. The mobile application risk identification method according to claim 1, characterized in that, The steps of inputting at least one detailed screenshot into the object detection model and outputting the category label, location information, and confidence score of the candidate object include: Each detail screenshot is input into the target detection model for forward propagation inference. The target detection model extracts features from each detail screenshot and generates a feature map. Based on the feature map, bounding box regression and classification prediction are performed on the preset anchor boxes. The category label, location information, and confidence score of each candidate target are output. The location information includes the bounding box coordinates, and the confidence score is used to characterize the probability that the candidate target belongs to the category indicated by the category label.
3. The mobile application risk identification method according to claim 1, characterized in that, The steps of cropping the target region where the candidate target is located based on the location information and extracting the visual feature vector of the target region include: Obtain the bounding box coordinates corresponding to the candidate target from the location information, crop the target region from the at least one detail screenshot based on the bounding box coordinates, input the target region into the feature extraction network, perform convolution and pooling operations on the target region through the feature extraction network, and output the normalized visual feature vector.
4. The mobile application risk identification method according to claim 1, characterized in that, The step of calculating the similarity distance between the visual feature vector and each vector in the camouflage image vector library includes: Each of the input vectors is read sequentially from the camouflage image vector library. The cosine distance or Euclidean distance between the visual feature vector and the currently read input vector is calculated. The calculated cosine distance or Euclidean distance value is used as the similarity distance. Wherein, the visual feature vector and the input vector are both normalized vectors of the same dimension.
5. The mobile application risk identification method according to claim 1, characterized in that, The step of determining the risk level of the mobile application under test based on whether there are candidate targets whose category labels belong to the risk identifier category and whether there are candidate targets that match the disguised image in at least one of the detail screenshots includes: Iterate through all candidate targets for each detail screenshot, determine whether the category label of each candidate target belongs to the risk identifier category, and determine whether each candidate target matches the disguised image; If any screenshot of the details contains a candidate target whose category tag belongs to the risk identifier category and a candidate target that matches the disguised image, the mobile application under test is directly determined to be of a high-risk level. When each screenshot contains only candidate targets whose category tags belong to the risk identifier category and no candidate targets that match the disguised image, the risk identifier score of each candidate target is accumulated. When the accumulated result exceeds the third threshold, the mobile application under test is determined to be of high risk level. Each candidate target has a risk identifier score set according to the risk identifier category corresponding to the category tag. When each screenshot contains only candidate targets that match the disguised image and no candidate targets whose category tags belong to the risk identifier category, the image risk score of each candidate target is accumulated. When the accumulated result exceeds the third threshold, the mobile application under test is determined to be of high risk level; otherwise, the mobile application under test is determined to be of low risk level.
6. The mobile application risk identification method according to claim 5, characterized in that, Based on the risk level, the steps for performing blocking or allowing the mobile application under test to be listed on the app store include: When the mobile application under test is determined to be of high risk, it will be blocked; when the mobile application under test is determined to be of low risk, it will be allowed to be listed on the app store.
7. The mobile application risk identification method according to claim 6, characterized in that, Also includes: Obtain the icon image of the mobile application to be tested, compare the similarity of the icon image with each vector in the dummy image vector library, and trigger a risk warning when the similarity exceeds the fourth threshold. When the mobile application under test is determined to be of a high-risk level based on the detailed screenshot and the risk warning is triggered, the risk warning is output as additional risk evidence along with the high-risk level. When the mobile application under test is determined to be of low risk level based on the detailed screenshot but triggers the risk warning, the risk level of the mobile application under test is adjusted to medium risk level and sent for manual review.
8. A mobile application risk identification system, characterized in that, include: The acquisition module is used to acquire at least one detailed screenshot of the mobile application under test; The detection module is used to input the at least one detailed screenshot into the target detection model and output the category label, location information and confidence score of the candidate target, wherein the category label includes a preset risk identification category; The extraction module is used to crop the target region where the candidate target is located based on the location information and extract the visual feature vector of the target region when the confidence score exceeds a first threshold. The comparison module is used to calculate the similarity distance between the visual feature vector and each vector in the camouflage image vector library, compare the similarity distance with a second threshold, and determine that the candidate target hits the camouflage image when the similarity distance is less than the second threshold. The determination module is used to determine the risk level of the mobile application under test based on whether there are candidate targets whose category tags belong to the risk identifier category and whether there are candidate targets that match the disguised image in the at least one detail screenshot; The execution module is used to perform interception processing or allow listing processing on the mobile application under test according to the risk level.
9. An electronic device, characterized in that, The electronic device includes a processor coupled to a memory storing at least one computer program, which is loaded and executed by the processor to enable the electronic device to implement the mobile application risk identification method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which, when executed by a processor, implements the mobile application risk identification method as described in any one of claims 1 to 7.
