Proxy IP optimization method, proxy IP optimization device and program product

Abstract

The invention provides a proxy IP optimization method, a proxy IP optimization device and a program product. According to one embodiment, the proxy IP optimization method comprises the following iteratively executed steps: a network environment sensing step: detecting and outputting network environment characteristics including traffic characteristics of a spatial dimension and traffic characteristics of a time dimension according to an access request and a reaction of a target network; a camouflage induction step: generating a camouflage strategy template for camouflage as normal network access and an induction strategy template for inducing a target network based on the network environment characteristics; and a proxy pool optimization step: generating network access traffic according to the disguise strategy template and the induction strategy template and sending the network access traffic to a target network.

Description

Technical Field

[0001] This invention mainly relates to the fields of information technology and network security, and in particular to methods, devices and programs for optimizing proxy IPs. Background Technology

[0002] In today's internet age, proxy IPs are used in many application scenarios (such as data quality monitoring, peer-to-peer traffic measurement, advertising compliance verification, security attack and defense evaluation, and compliance sampling access). These proxy IPs use virtual IP addresses to access websites, simulating normal, large-scale network access behavior. Additionally, proxy IPs are sometimes used for network security reasons, preventing the other party from recording one's real IP address. This effectively prevents sensitive data such as personal information and geolocation from being tracked or leaked, reducing the risk of cyberattacks.

[0003] On the other hand, websites also detect whether a visitor is using a proxy IP and may block access if a proxy IP is detected. To prevent detection, the industry has proposed several proxy IP protection strategies. For example,

[0004] Frequent IP switching strategy: Switch proxy IPs according to fixed time intervals or access frequency thresholds, thereby avoiding rapid blocking caused by repeated use of a single IP.

[0005] Static proxy pool management: A large number of proxy nodes are built into a static pool and allocated in a random or round-robin manner.

[0006] Simple traffic spoofing: Randomize or template HTTP headers (User-Agent, Accept-Language, Referer, etc.).

[0007] One-way "detection → generation" or "generation → scheduling" chain: Some systems use the detection results for the generation module, or directly hand over the generation strategy to the scheduler for execution.

[0008] Meanwhile, the website is continuously upgrading its proxy IP detection technology. Currently, the website's detection capabilities have evolved from simple frequency control and rule auditing to multimodal and multi-layered comprehensive detection. For example, it employs detection technologies including request frequency / rate analysis, behavioral sequence modeling, HTTP / TLS fingerprinting, encrypted handshake feature comparison, latency and congestion profiling, and group similarity clustering.

[0009] Faced with constantly evolving proxy IP detection technologies, traditional proxy IP protection strategies sometimes fall short. For example, while frequent IP switching strategies are simple in structure and easy to implement, their switching rhythm and patterns can be identified through time-series analysis, and they cannot handle cross-layer fingerprinting and behavioral modeling. Static proxy pool management strategies can distribute access sources, but they lack dynamic optimization based on environmental feedback, resulting in low IP resource utilization, untimely isolation of abnormal nodes, and high costs. Simple traffic masquerading strategies can enhance surface masquerading, but they do not adequately cover deep features such as TLS / JA3 fingerprints, request timing, and cross-session semantic consistency, making them difficult to deal with deep detection. One-way "detection → generation" or "generation → scheduling" chain strategies, lacking reverse feedback and multi-objective unified optimization, suffer from slow overall adaptive speed, poor global optimality, and slow response to rule mutations.

[0010] Therefore, the industry needs a systematic solution that can adaptively evolve, actively disrupt the detection statistical space, and optimize resources under the constraints of service level agreements (SLAs). Summary of the Invention

[0011] The inventors of this invention have proposed this invention in view of the problems in the aforementioned background art. In this invention, an adaptive closed-loop mechanism of environmental perception—camouflage and deception—resource scheduling is proposed to achieve long-term stable availability of proxy IPs under frequent changes in unknown detection strategies.

[0012] In one aspect, the present invention provides a proxy IP optimization method, comprising the following iteratively executed steps: a network environment awareness step, which detects and outputs network environment features including spatial and temporal traffic features based on the access request and the response of the target network; a masquerade and inducement step, which generates a masquerade strategy template for masquerading as normal network access and an inducement strategy template for inducing the target network based on the network environment features; and a proxy pool optimization step, which generates network access traffic and sends it to the target network based on the masquerade strategy template and the inducement strategy template.

[0013] The proxy IP optimization method of the present invention further includes the following steps: a risk energy detection step, which detects the risk energy or the rate of change of risk energy representing the failure risk of the proxy IP; and a masquerade-induced adjustment step, which adjusts the strategy exploration level and template perturbation amplitude in the masquerade-induced step when the detected risk energy or the rate of change of risk energy exceeds a preset threshold.

[0014] In the proxy IP optimization method of the present invention, when the risk energy or the rate of change of risk energy is detected to exceed a preset threshold, the degree of strategy exploration is increased, and the perturbation amplitude of the spoofing strategy template and / or the inducement strategy template is increased.

[0015] As an example, in the network environment perception step, convolutional neural networks are used to extract spatial traffic features, and long short-term memory networks are used to extract temporal traffic features.

[0016] As an example, in the proxy pool optimization step, access requests that mimic the traffic characteristics of normal users are generated based on the masquerade strategy template.

[0017] As an example, in the proxy pool optimization step, access requests are generated based on the inducement strategy template to interfere with the detection model of the target website.

[0018] As an example, in the network environment perception step, the output is a network environment feature vector containing traffic features in both spatial and temporal dimensions; in the spoofing and inducement step, a generative adversarial network is used, taking the network environment feature vector and the features representing the network response of the object as inputs to the generative adversarial network, and generating an inducement policy template based on the output of the generative network in the generative adversarial network.

[0019] In another aspect, the present invention also provides a proxy IP optimization device, comprising: a network environment awareness module, which detects and outputs network environment features including spatial and temporal traffic features based on access requests and the response of the target network; a masquerade and inducement module, which generates a masquerade strategy template for masquerading as normal network access and an inducement strategy template for inducing the target network based on the network environment features; and a proxy pool optimization module, which generates network access traffic and sends it to the target network based on the masquerade strategy template and the inducement strategy template.

[0020] In different embodiments, the proxy IP optimization apparatus of the present invention can be configured to perform various processes of the aforementioned proxy IP optimization method.

[0021] As an example, the proxy IP optimization device of the present invention may include a risk energy detection module, which detects the risk energy or the rate of change of risk energy that represents the failure risk of the proxy IP. When the risk energy detection module detects that the risk energy or the rate of change of risk energy exceeds a preset threshold, the masquerade induction module changes the strategy exploration level and the template perturbation amplitude.

[0022] In another aspect, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the proxy IP optimization method described above.

[0023] In another aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned proxy IP optimization method.

[0024] In another aspect, the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the aforementioned proxy IP optimization method. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below only relate to some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0026] Figure 1 This is a schematic flowchart of a proxy IP optimization method according to an embodiment of the present invention.

[0027] Figure 2 This is a schematic flowchart of a proxy IP optimization method according to another embodiment of the present invention.

[0028] Figure 3 This is a schematic structural block diagram of a proxy IP optimization device according to an embodiment of the present invention.

[0029] Figure 4 This is a schematic structural block diagram of a proxy IP optimization device according to another embodiment of the present invention.

[0030] Figure 5 This is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, some embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be noted that the illustrative embodiments and descriptions described herein are for explaining the present invention, but are not intended to limit the present invention.

[0032] Furthermore, the acquisition, storage, use, and processing of data described in this manual comply with the relevant provisions of national laws and regulations.

[0033] It should be noted that certain software, components, models, and other existing solutions in the industry may be mentioned in this specification. These should be considered as exemplary and are intended only to illustrate the feasibility of implementing the technical solution of this application. However, they do not mean that the applicant has used or necessarily used such solutions.

[0034] Implementation Method 1

[0035] Figure 1A schematic flowchart of a proxy IP optimization method according to one embodiment of the present invention is shown. Figure 1 As shown, the proxy IP optimization method includes a real-time network environment awareness step S1, a masquerade and inducement step S2, and a proxy pool optimization step S3. In the network environment awareness step S1, based on the access request and the target network's response, network environment features containing both spatial and temporal traffic characteristics are detected and output. This serves as the input basis for the masquerade and inducement step S2. In the masquerade and inducement step S2, a masquerade strategy template for masquerading as normal network access and an inducement strategy template for inducing access to the target network are generated based on the awareness feedback from step S1. In the proxy pool optimization step S3, network access traffic is concurrently sent to the target network according to the masquerade strategy template and the inducement strategy template. Thus, these three steps form a closed-loop system, continuously adapting to the target network's detection strategy through iterative iteration to achieve the best proxy IP protection effect.

[0036] The following provides further exemplary details of each step.

[0037] When the target network employs anti-scraping techniques, the key to these techniques lies in their adaptability and accuracy. For example, the target network might adjust its detection strategy based on the following characteristics:

[0038] • Access frequency: Detects proxy IPs based on access frequency. For example, if the access frequency of a certain IP exceeds a preset threshold, it is determined to be a proxy IP and access is restricted.

[0039] • Traffic distribution: Detect proxy IPs based on traffic distribution. For example, if the access interval from a certain IP is very fixed, it is judged to be a proxy IP and access is restricted.

[0040] • HTTP header consistency: Some proxies may not be able to correctly forge certain HTTP header fields, so some detection strategies will compare the HTTP headers in the access request. If an inconsistency is found, it will be judged as a proxy IP and access will be restricted.

[0041] Therefore, relying solely on fixed masquerade strategies in existing technologies makes proxy IPs prone to failure. Furthermore, once the target network upgrades its detection policies, the probability of the proxy IP being blocked increases, raising the risk of service interruption. In this invention, dynamically detecting the environmental characteristics of the target network is key to achieving effective proxy IP protection.

[0042] To achieve real-time environmental detection, in one embodiment of the present invention, data collection is first performed in the environmental perception step S1 described above. For example, request / response logs, HTTP header sequences, TLS handshake elements, RTT, request interval, session duration, region, task identifier, and proxy node ID can be collected from traffic data.

[0043] The following is an example of a data structure collected in one embodiment of the present invention.

[0044]

[0045] Next, as an example, we will use a multi-layer architecture based on "convolutional neural network (CNN) and long short-term memory network (LSTM)" to extract traffic features, that is, to extract traffic features in the spatial dimension and traffic features in the temporal dimension respectively.

[0046] For spatial traffic characteristics, as a concrete example, a two-dimensional traffic characteristic matrix can be constructed for each request. The resulting matrix M is in the form shown below.

[0047]

[0048] More specifically, the extracted traffic packet features can come from fields in RequestLog, including but not limited to: UA, Accept-Language, Header sequence and length, TLS handshake parameters (Cipher Suites, Extensions), RTT value, previous access interval, session duration, HTTP status code, response byte count, etc. Thus, each row in matrix M corresponds to a specific field category, and each column corresponds to the numerical or embedding vector representation of that field, such as one-hot encoding, hash embedding, or numerical normalization, thereby achieving a structured representation of traffic features.

[0049] In addition, constructing heterogeneous graphs ,in:

[0050] Node V: Fields / Parameters / Statistics;

[0051] side Co-occurrence relationships, dependency relationships, and temporal adjacency relationships (reflecting cross-layer semantic connections).

[0052] More specifically, the nodes of the heterogeneous graph G are derived from the fields / parameters / statistics extracted from the traffic packets mentioned above, and the edge relationships are constructed based on the co-occurrence relationship between fields in the traffic packets, the contextual dependency relationship (such as the continuous occurrence of a combination of a certain TLS parameter and a specific UA), and the temporal adjacency relationship, in order to reflect cross-layer semantic associations.

[0053] In a specific example, for HTTP headers, "User-Agent," "Accept-Language," "Cookie field length," and "Header field sequence order" can be used as node features. For TLS, "Cipher Suites set," "Extensions list length," and "JA3 fingerprint hash" can be used as node features. For the time dimension, "last request interval," "RTT fluctuation," and "session duration" can be used as node features. If the above fields co-occur multiple times within the same time window, an edge relationship can be established.

[0054] Next, the feature map is obtained through the convolutional layers of the Convolutional Neural Network (CNN).

[0055]

[0056] in, This represents the k-th feature map. Let be the k-th convolution kernel, * denotes convolution operation, σ denotes activation function, and M denotes the aforementioned traffic feature matrix.

[0057] In addition, as shown in the following equation, the heterogeneous graph G obtained above is processed by a graph attention network (GAT) neural network.

[0058]

[0059] here, It is the new feature representation of node i after attention aggregation, which is the result of weighted summation of the features of its neighboring nodes. Let i represent the set of neighbors of node i. is the attention coefficient, used to measure the importance of node j's features to node i. After normalization using the softmax function, it ensures that the sum of the weights of all neighbors is 1. W is a trainable weight matrix used to adjust the original features of nodes i and j. and Linear transformations can be performed to increase the dimensionality of features or compress information.

[0060] This indicates that the scores of all neighboring nodes j are normalized to ensure... This gives the weights an interpretable “attention share” meaning. This is the activation function, used to introduce nonlinearity into scalar fractions. This indicates a feature concatenation operation, which will... and The feature vectors are concatenated into a longer vector. It is a trainable attention vector, which is inner-producted with the concatenated feature vector to obtain a scalar score, used to measure the strength of the association between i and j.

[0061] Thus, through the linear transformation of the neighbor node features, attention-weighted summation, and normalization processes described above, a new feature vector for each node i can be obtained. This feature vector is used to express the representation ability of the node after cross-field and cross-layer semantic fusion.

[0062] Next, the obtained feature map Pooling is performed on the heterogeneous graph G described above, and readout processing is applied, such as global average pooling or global weighted pooling. The results of these two operations are then concatenated to obtain a vector. .

[0063]

[0064] In this way, the flow characteristics in the spatial dimension were obtained.

[0065] In addition, in step S1, a Long Short-Term Memory (LSTM) network can be used to extract the temporal patterns of access behavior.

[0066] Therefore, feature vectors of time series are constructed.

[0067]

[0068] This represents the nth feature of the current time step, such as access intervals, state distributions, RTT fluctuations, session cycles, etc., extracted from traffic data according to a pre-set time period.

[0069] The operation is performed using a Long Short-Term Memory (LSTM) network as shown below.

[0070]

[0071] here, The above-mentioned feature vector is the input at the current time step. This represents the hidden state of the previous time step, which is the intermediate feature representation of the LSTM output after processing the (t-1)th input, containing the memory of historical information. This represents the hidden state at the current time step. This represents the output at the current time step. In other words, at step t, the LSTM calculates the current memory and the current output based on the current input and the memory from the previous step. Thus, by continuously updating the memory and progressively processing the input sequence, new memories and outputs are dynamically generated based on historical information and the current input, thereby capturing long-term sequence dependencies in traffic data and obtaining traffic features in the time dimension.

[0072] Specifically, as an example, the time-series feature vector sequence arranged in chronological order is first... The input is fed into a temporal convolutional network, which extracts long-term and short-term dependencies through multiple layers of causal convolution and dilated convolution. The temporal convolutional network outputs the corresponding temporal features at each time step, and then performs temporal pooling on the sequence, such as average pooling or weighted pooling. After temporal pooling of the traffic features in the temporal dimension, they are concatenated and fused with the feature vectors in the spatial dimension, as shown in the formula below.

[0073]

[0074] Here, Concat indicates concatenation. This represents the combined feature vector obtained by concatenation. These are the spatial dimension feature vectors obtained earlier, and TemporalPool represents the temporal pooling operation. It is a time-series characteristic sequence.

[0075] Subsequently, feature transformation and activation operations are performed as shown in the following equation to fuse the features. Linear transformation and nonlinear activation are performed to obtain the final output features. .

[0076]

[0077] here This represents a non-linear activation function, such as ReLU, Sigmoid, Tanh, etc., used to introduce non-linearity and enhance the expressive power of the model. These are the weight matrix and the bias term, used to adjust the aforementioned eigenvectors. Perform a linear transformation. It is a bias term that can be continuously iterated and optimized during training.

[0078] Through such feature transformation and activation, a vector representing environmental risk is obtained. Such feature vectors It includes the temporal and spatial characteristics of the network environment in which the proxy IP is located, and can comprehensively reflect risk indicators that could lead to the proxy IP being blocked by the target network, such as access interval variance, header consistency, TLS mode complexity, state anomaly ratio, and latency fluctuation.

[0079] The eigenvectors were obtained as described above. Then, the process proceeds to step S2, where a policy template for the proxy IP is generated based on environmental characteristics.

[0080] In step S2, a corresponding spoofing strategy template is generated based on the perception results obtained in step S1. Here, "template" refers to the standardized format for configuring or calling a proxy IP and a set of corresponding parameters. The "spoofing strategy template" is a standardized strategy used to spoof traffic characteristics (such as HTTP fields). It reduces the identifiability of the proxy IP by adjusting traffic characteristics (such as request headers and protocol fields), making the traffic more similar in form to the traffic characteristics of a normal user, thus causing the target website to judge it as normal user access. Furthermore, the inventors also discovered that merely spoofing traffic characteristics (such as HTTP fields) cannot completely conceal the existence of a proxy IP. It is possible to effectively confuse the target network's detection system by generating inducement traffic that is highly similar to real user behavior, preventing it from detecting access from a proxy IP. Therefore, in step S2, a so-called "inducement strategy template" for injecting inducement traffic to confuse the target website is also generated. Here, the "inducement strategy template" is used to generate access characteristics that are consistent with or highly similar to the statistical distribution of real users, making it impossible for the detection model on the target website to determine whether the access source is a proxy IP from the observable traffic, thereby reducing the probability of being blocked and extending the life cycle of proxy nodes.

[0081] Specifically, this can be based on the environmental feature vector generated in the previous step S1. This drives the generation of policy templates and parameter adjustments. For example, the environmental feature vector can be determined as shown above. Then, based on environmental feature vectors Given the current task context (e.g., task priority, geographic policy, security compliance flags, etc.), output a set of candidate values ​​for template parameters. For example, a trainable mapping network (e.g., a shallow mapper or regression network) or a rule-based lookup mechanism can be used to implement the mapping, incorporating confidence metrics and business constraint verification during the mapping process. To ensure system controllability and auditability, the mapping output must pass through a template compliance checker (including SLA verification, privacy / compliance rule verification, and rollback condition generation) before becoming the final template version. The above mapping, verification, and versioning operations can be recorded in the audit log, including input... The summary, mapping module version number, compliance verification results, and template identifier are used to facilitate backtracking and supervision.

[0082] As a specific example, environmental feature vectors can be used... The feature vector is concatenated with the task context vector to form a total feature vector. Furthermore, outlier filtering can be performed to remove unreasonable values ​​from the total feature vector, such as correcting load values ​​greater than 100% to 100%. Feature enhancement can also be performed, applying cross-feature processing (e.g., "risk level × concurrent access volume") to key features (such as the risk level of the accessed area) to improve mapping accuracy. Dimension alignment can also be performed to ensure that the dimension of this feature vector matches the input dimension of subsequent mapping modules.

[0083] As an example, a lightweight neural network, such as a shallow network, can be used to map the processed feature vectors to a policy template. Specifically, historical environmental features ( Historical values ​​and corresponding optimal template parameters (which can be manually labeled or effective parameters determined based on business metrics) are used as training data. A loss function is designed according to the parameter type; for example, the MSE loss function is used for numerical parameters, and the cross-entropy loss function is used for categorical parameters. The model is then trained in this way. Additionally, a confidence metric can be introduced into the model, adding a "confidence branch" to the output layer to measure the reliability of the current mapping result. For example, a confidence score between 0 and 1 can be output for each parameter value; the higher the score, the more reliable the result. During inference, the preprocessed total feature vector is input, and the network outputs parameter values ​​and their corresponding confidence scores. Parameter values ​​with confidence scores greater than a threshold can be selected to form a candidate parameter set, which is the template mentioned earlier.

[0084] As another example, the aforementioned total feature vector can be mapped to a policy template using a rule-based lookup table mechanism. The advantage of this is that it requires no training and facilitates manual intervention and auditing. For example, a mapping table illustrating the correspondence between feature vectors and parameter values ​​can be prepared in advance. Below is a schematic example of such a mapping table.

[0085]

[0086] When looking up a table, conditional matching can be performed on the input feature vector. For example, a load of "75%" can match the rule "≥70%". If multiple rules are matched, several predetermined parameter values ​​are selected as candidates in descending order of confidence. If no rule is matched, for example, the default parameter can be triggered and the confidence level can be set to 0.5.

[0087] Regardless of the mapping method used, business constraint validation must be performed when generating candidate parameters to eliminate obviously unreasonable parameters. As an example, the following constraint rules could be used:

[0088] Request frequency ≤ target server's maximum capacity frequency (from (derived from "server load").

[0089] The number of concurrent connections is less than or equal to the maximum concurrency allowed by the network environment (e.g., ≤10 for public networks, ≤50 for private networks).

[0090] The privacy encryption switch must be consistent with the "compliance flag" of the context (flag = yes → switch = on).

[0091] If a candidate parameter violates the constraints, it can be directly eliminated and replaced with a suboptimal candidate value. For example, assuming the request frequency of the mapped network output is 30 times / second, which violates the constraint "public network ≤ 10", it can be replaced with 10 times / second, and the confidence level can be reduced by 0.1.

[0092] Next, compliance checks can be performed on the candidate parameter set. As an example, the following items could be included.

[0093] SLA Validation: Template parameters must meet the Service Level Agreement (SLA). For example, if the request success rate is ≥99.5%, then the "error request tolerance" must be ≤0.5%.

[0094] Privacy / Compliance Rule Verification: If the context has a "Privacy Encryption" flag, verify whether the "Encryption Switch" is on and whether the encryption algorithm complies with GDPR / security level protection requirements;

[0095] Geographic policy validation: Verifies whether the "access region" is on the allowed list. For example, if overseas access is not allowed, then overseas User-Agents will be removed.

[0096] Rollback condition generation: Add "abnormal rollback rules" to the template. For example, if the spoofing success rate is ≤80%, automatically roll back to the previous version of the template.

[0097] If the validation passes, the candidate parameter set can be used as the final template parameters to generate the final policy template. If the validation fails, return to the previous mapping stage, adjust the candidate parameters, and regenerate and validate again.

[0098] After the strategy templates are generated, version tagging and log retention can be performed for backtracking. For example, a unique identifier can be assigned to each template, which can include the version number, generation time, and mapped module version.

[0099] The parameter set (strategy template) generated in this way constitutes the current masquerade strategy. Request traffic is generated according to the masquerade strategy. This traffic is disguised as the access behavior of normal users, so it is also called masquerade traffic.

[0100] Additionally, in step S2, Generative Adversarial Networks (GANs) can be used to generate inducement strategy templates to generate traffic similar to real user behavior, i.e., the induced traffic mentioned earlier, to interfere with the target website's detection model. In other words, the purpose of generating so-called "induced traffic" is to cause the target website's detection system (such as anti-crawler, abnormal access identification, malicious behavior interception models, etc.) to make judgment biases by using deceptively realistic induced traffic. For example, it can reduce the accuracy of the detection model, making it unable to distinguish between real user traffic and induced traffic, and mistakenly classifying induced traffic as legitimate traffic. It can also affect the model's feature learning logic. If the detection model relies on user behavior features (such as access frequency, operation path, device fingerprint, etc.) for training, continuous induced traffic will interfere with the model's learning of real user features, leading to a continuous decline in the model's subsequent recognition ability.

[0101] Generative Adversarial Networks (GANs), also known as Game Adversarial Networks, consist of a generator network G and a discriminator network D. The technical concept of GANs originates from zero-sum games in game theory. The generator is responsible for generating data samples, while the discriminator is responsible for determining whether the input samples are real data or fake data generated by the generator. During training, the generator and discriminator train together through a game-like interaction. The generator strives to generate more realistic samples to induce the discriminator, while the discriminator continuously improves its discrimination ability. The final training result is that the data generated by the generator becomes increasingly realistic, and the discriminator's discrimination ability correspondingly increases, until a Nash equilibrium is reached.

[0102] The training of a generative adversarial network in one embodiment of the present invention will be described below.

[0103] In this embodiment, to ensure that the generator G output in the generative adversarial network is highly consistent with the real user behavior of the target site in terms of statistics and spatiotemporal dimensions, an environmental feature vector is used. The training process for Conditional Adversarial Generation (cGAN / TimeGAN) as conditional inputs can specifically include the following steps:

[0104] 1) Training Data Construction: As described above, traffic data (RequestLog) can be collected by task dimension and segmented by time window (e.g., 10-minute window). For each request, calculate and store its spatial feature matrix M (field embedding / hashing / normalization) and temporal feature vector T (interval statistics, RTT distribution, session duration, etc.). Requests within the same task / region / time window are stratified by quality (low anomaly, high normality) to form a candidate sample pool.

[0105] 2) Pseudo-labeling and semi-supervised updating: Online metrics (online ban rate, 429 / 403 trigger rate, challenge / CAPTCHA trigger rate, session anomaly rate, etc.) are used to perform semi-supervised pseudo-labeling on candidate samples. Specifically, for example, let... Indicates "approximately real user". This indicates "suspected of being detected / abnormal." The pseudo-label threshold uses a sliding window statistical approach (e.g., if an IP has less than 1% of 429 / 403 events within the window and a request success rate greater than 95%). ;otherwise Pseudo-labels also introduce confidence weights. , used for training loss weighting.

[0106] 3) Condition Construction: Constructing environmental feature vectors The sample M and T are concatenated and used as the conditional input to the generator G. The discriminator D takes the generated sample or the real sample as input and the corresponding sample. .

[0107] 4) Training objective: Employing joint loss

[0108]

[0109] in To combat losses under normal conditions, To impose mutual information constraints (which can be approximated using MINE or InfoNCE upper bounds) to minimize the mutual information I(S;Z) between the agent identity S and the observable features Z, This is a spatiotemporal consistency loss (e.g., the norm of the feature differences between adjacent time slices or DTW). The supervision loss is weighted by the confidence level for false labels. Adjust the various hyperparameters. Then, adjustments can be made during project implementation using cross-validation or online A / B testing.

[0110] 5) Training Process and Online Updates: Initially, G and D are trained offline (using historical high-confidence samples). After deployment, small-batch online fine-tuning is employed: at each time step t, new samples and pseudo-labels are collected within window W, and sampling is performed using a priority-based empirical replay mechanism (prioritizing hard examples). D is updated with a small learning rate, and then G is updated using the discriminant feedback of D (alternating training). Statistical descriptions of the generated samples (Header distribution, JA3 clusters, interval distribution) can also be recorded during training for online comparison.

[0111] Through iterative training, once the generator and discriminator reach a balance, the generator's output can be mapped to the following template set:

[0112] • header_tpl: HTTP header field sequence and discrete perturbation range (e.g., field order, Accept-Language distribution, Referer pattern);

[0113] • tls_tpl: TLS / JA3 variant cluster identifier and subset of handshake parameters;

[0114] • interval_tpl: Baseline interval and jitter distribution (e.g., base=700ms, jitter ~ Normal(0,120ms));

[0115] •disguise_level: Overall perturbation strength (value 0–1, subject to SLA upper limit).

[0116] Each template can include a TTL, version number, and statistical description for rollback and auditing. Furthermore, the mapping here uses predefined parsing rules (such as linear transformations, table lookups, cluster matching, etc.) to convert the vectors generated by generator G into structured templates that conform to network protocol specifications and can be used for traffic generation. The spoofing effect and stability are balanced through disguise_level and SLA constraints.

[0117] After generating the spoofing strategy template and the inducement strategy template, the proxy IP optimization method in this embodiment proceeds to step S3 to optimize the proxy pool.

[0118] The proxy pool mentioned here refers to a collection of various resources that can be used to proxy IPs. For example, it may include a collection of proxy IPs and a combination of policy template parameters bound to each proxy IP, such as header perturbation magnitude, TLS parameter selection range, time interval jitter level, etc.

[0119] For example, the inducement policy template described above can be serialized together with the spoofing policy template into JSON / Protobuf, and then pushed to the SDK / proxy gateway by the policy delivery and return layer. An example payload is shown below:

[0120] {

[0121] "proxy_id":"p-9012",

[0122] "header_tpl":"...",

[0123] "tls_profile":"ja3-xxxx",

[0124] "interval_policy":{"base_ms":700,"jitter_ms":120},

[0125] "disguise_level":0.6,

[0126] "ttl":3600

[0127] }

[0128] In addition, the scheduling layer can perform SLA verification on the template before deployment, such as ensuring that the success rate of critical tasks is ≥X% and the average latency is ≤Y ms. If the template does not meet the SLA, the dispatch level will be automatically reduced or the template will be switched to a low-disturbance sub-template.

[0129] Additionally, an anonymity score can be calculated for each proxy IP based on spoofing logs and inducement effects. This anonymity score quantifies the anonymity performance of each proxy node within the current period and can be constructed using weighted scores based on metrics such as spoofing effect logs, induced traffic feedback, and access success rate. The following formula is an example of how to calculate the anonymity score. In the formula, w1, w2, and w3 represent the weights of different anonymity metrics, which can be preset based on historical data or experience.

[0130] S = w1·s1 + w2·s2 + w3·s3

[0131] in:

[0132] s1 is the concealment score obtained based on the camouflage effect log, which is used to characterize whether the target site's blocking rate, challenge / CAPTCHA trigger rate, HTTP / TLS fingerprint similarity, etc., are close to the real user distribution under the current camouflage strategy;

[0133] s2 is the sub-score of the inducement effect obtained based on the induced traffic feedback. It is used to characterize the degree to which the generated traffic is misjudged as real traffic by the target detection model or absorbed into the other party's training / verification process, such as the proportion of induced samples accepted and the proportion of induced false alarms.

[0134] S3 is an availability sub-score based on access success rate and stability, used to reflect the success rate of requests, retry rate, average latency fluctuation and other business availability indicators of the proxy node in the current period.

[0135] s1, s2, and s3 can be obtained by normalizing, weighting, or scoring the original indicators mentioned above.

[0136] Resource allocation adjustment: Based on the anonymity score, a reinforcement learning model is used to adjust the allocation strategy of proxy pool resources. For example, high-anonymity IPs are prioritized for critical tasks, while low-anonymity IPs are placed in reserve.

[0137] In this way, the IP with the best current concealment effect (i.e., disguise effect) can be selected from the proxy resource pool, and access traffic and inducement traffic sent to the target network can be generated according to the disguise strategy template and inducement strategy template issued in step S2, respectively.

[0138] For example, the template issued in step S2 can be first parsed in a structured manner to extract all executable parameters and categorize them into configuration items that can be directly called. Below is an example parsing result.

[0139]

[0140] For example, the frequency, interval, and concurrency of requests can be controlled according to the traffic sending rules defined in the template. Taking the table above as an example, based on "request frequency = 8 times / second", the basic request interval is calculated as 1000ms / 8 = 125ms. This is compounded by "request interval fluctuation = 50-200ms", with a random fluctuation value generated before each request, such as 125ms + 80ms = 205ms or 125ms - 30ms = 95ms. Furthermore, based on "concurrent connections = 8", the number of requests initiated simultaneously by a single IP can be limited using a thread pool / coroutine pool; for example, the maximum number of threads can be limited to 8.

[0141] Next, we construct the request headers. The request headers are crucial for identifying traffic identity; they must be constructed according to template rules to ensure consistency with normal user request header fields, avoiding identification as crawlers / abnormal traffic. Taking the table above as an example, it can include the following content.

[0142] User-Agent selection: Randomly select a real UA from the "PC Chrome 120.0" pool specified in the template, such as Mozilla / 5.0 (Windows NT 10.0; Win64; x64) AppleWebKit / 537.36 (KHTML, likeGecko) Chrome / 120.0.0.0 Safari / 537.36).

[0143] Referer generation: According to the "random within the same domain" rule, first obtain the target domain (such as https: / / target.com), and then randomly concatenate the sub-paths under that domain (such as https: / / target.com / home, https: / / target.com / category / phone).

[0144] Cookie carrying: Generate a session cookie that conforms to the HTTP standard (e.g., sessionid=abc123xyz;expires=Thu, 20-Nov-2025 16:00:00 GMT; path= / ), and carry it in subsequent requests until it expires;

[0145] Additional fields: In addition to the fields Accept (content type), Accept-Encoding (encoding method), Connection: keep-alive, etc., as per normal browser requests, to avoid missing fields.

[0146] In addition, user access behavior can be defined based on policy templates. Taking the table above as an example, the following content can be included:

[0147] Access path generation: Adapted to "click path randomness = 0.7", where 0.7 represents "70% random + 30% fixed". First, the core access path of the target network is preset (e.g., homepage → category page → product page → shopping cart page). Each time accessed, there is a 30% probability of following the fixed path, and a 70% probability of randomly inserting / skipping some pages within the fixed path (e.g., homepage → product page → category page → shopping cart page). An example path could be: https: / / target.com → https: / / target.com / category / phone → https: / / target.com / product / 123 → https: / / target.com / cart;

[0148] Page Dwell Time and Action Simulation: Adapt to "page dwell time = 5-10 seconds". After each page visit (i.e., after receiving a response), pause for 5-10 seconds randomly to simulate real user browsing content before initiating the next request. If the page contains interactive elements (such as buttons or forms), click / input behaviors can be simulated with a low probability (e.g., 10%), such as submitting random but valid content to a form, avoiding empty submissions.

[0149] Error Request Control: Adapted to "Error Request Tolerance = 0.3%", normal users will occasionally trigger errors such as 404 (accessing a non-existent page) and 400 (incorrect request parameters). Error requests will be randomly generated at a rate of 0.3% (such as accessing https: / / target.com / product / 99999 (non-existent ID) or submitting incorrectly formatted parameters), but the error rate will be ensured not to exceed the threshold to avoid being judged as malicious scanning.

[0150] Additionally, privacy compliance processing is performed for the "Compliance Configuration Parameters". If the template requires privacy encryption, sensitive fields in the request (such as user ID, phone number, address, etc.) are encrypted to ensure that traffic complies with compliance requirements. For example, this could include the following:

[0151] Identify sensitive fields: Define sensitive fields according to business rules (such as user_id and phone in request parameters);

[0152] Encryption processing: Encrypt sensitive fields using the AES-256 algorithm specified in the template (e.g., user_id=12345 → encrypted as user_id=eFgH1234xyz);

[0153] Log reporting: Report the request summary (excluding sensitive information), IP identifier, template version and other information to the audit log at a frequency of "1 minute / time".

[0154] Therefore, as described above, the constructed request (including a disguised request header, simulated behavior path / parameters, etc.) is sent to the target network using the selected optimal hidden IP. Furthermore, the current IP's connection status (whether it is normal, remaining concurrency), generated cookie information (to ensure the continuity of subsequent requests), visited paths, dwell time, etc. (to avoid duplicate visits or contradictory behavior) can be recorded in the session context (such as local cache, database).

[0155] Additionally, if the request fails, such as encountering a 503 server unavailable error or a 403 block, it will be retried according to the template's implicit rules, for example, retrying twice with a 3-second interval. If it still fails, the proxy IP will be switched, meaning the optimal IP will be reselected from the resource pool.

[0156] Furthermore, in this invention, steps S1, S2, and S3 are performed iteratively. As described above, after generating network access traffic and sending it to the target website based on the spoofing strategy template and the inducement strategy template in step S3, the network environment detected in real time in step S1 may also be updated. Therefore, updated environmental features are generated in step S1, and updated spoofing strategy templates and inducement strategy templates are generated based on the updated environmental features in step S2. Then, network access traffic is generated based on the new spoofing strategy template and inducement strategy template in step S3. This iterative process continues, and the neural network model used in each step is continuously trained and optimized, ultimately generating access traffic in step S3 that sufficiently approximates real user access behavior.

[0157] Implementation Method 2

[0158] As another embodiment of the present invention, an inner-loop self-learning process can be added to the above-described embodiment 1. Specifically, the network environment sensing, generation of camouflage and inducement strategy templates, and optimization of proxy resource pools in embodiment 1 are iteratively arranged to form a closed loop, which can be called the "outer loop". In this embodiment, the model parameters are adjusted according to the network environment characteristics to adjust the intensity of camouflage and inducement, which also forms a closed loop, referred to here as the "inner loop". Figure 2 A schematic flowchart of a proxy IP optimization method according to an embodiment of the present invention is shown.

[0159] As a specific example, based on the proxy IP optimization method in Implementation Method 1 above, in... Figure 2 In step S4 shown, risk energy is defined. .

[0160] here, It is the environmental feature vector generated in the aforementioned network environment sensing step, also known as the risk vector, which can include multiple statistical features such as request interval fluctuations, abnormal proportion of status code distribution, header consistency deviation, and TLS handshake mode changes.

[0161] Risk vector The i-th risk statistical component in the data can be a numerical representation of factors such as the ban rate offset, the proportion of abnormal status codes, the header deviation value, and the TLS fingerprint difference.

[0162] The weights corresponding to the risk statistical components are used to reflect the degree of influence of the risk statistical components on the overall risk assessment. It can be preset based on historical operating data or experience.

[0163] Here we adopt This is to amplify the sensitivity of high deviation risk components, making risk energy more sensitive to abnormal risk components.

[0164] Alternatively, it can be defined. That is, the rate of change of global risk energy.

[0165] when or At that time, the inner loop self-learning is triggered.

[0166] here, This represents the risk threshold for triggering inner-loop self-learning. The threshold representing the rate of change of risk energy in the inner loop self-learning is used to identify risk mutation situations.

[0167] Here, the risk energy defined above reflects the comprehensive risk status of access behavior in the actual operation scenario in both the time dimension and the cross-layer semantic dimension. It can characterize the potential exposure risk of the current proxy access from the perspective of the target network detection. When the risk energy increases or the rate of change of the risk energy rises abnormally, it can be determined that there may be situations such as detection strategy upgrades, feature mutations, or increased blocking risks, thereby triggering inner-loop self-learning, i.e., the strategy readjustment mechanism.

[0168] The inner-loop self-learning mentioned here can be achieved by adjusting the two parameters β and ε.

[0169] β is the policy temperature coefficient, used to adjust the degree of policy exploration. When risk volatility increases, β is increased to explore more stochastic policies, while β is decreased when the environment is stable to tend towards the existing optimal policy and improve stability.

[0170] ε is a camouflage / induction strength parameter used to control the magnitude of perturbations in the template, such as header perturbation, TLS fingerprint variation range, and temporal jitter intensity. Increasing ε enhances the camouflage effect when the risk increases, and decreasing ε reduces overhead when the risk decreases.

[0171] Increasing the spoofing / inducement strength ε can be achieved by, for example, increasing the range of header randomization, increasing the variation of TLS handshake parameters, increasing the time jitter, and prioritizing the use of strongly spoofed templates. Conversely, decreasing the parameter ε reduces perturbation and maintains stability, thereby enabling high-speed access.

[0172] Third Implementation Method

[0173] The above description uses a proxy IP optimization method as an example to illustrate some specific embodiments of the present invention, but the present invention can also be implemented as a proxy IP optimization device. Figure 3 This is a schematic diagram of a proxy IP optimization device according to an embodiment of the present invention.

[0174] like Figure 3 As shown, the proxy IP optimization device includes a network environment awareness module 301, a masquerade inducement module 302, and a proxy pool optimization module 303. Each module can be implemented by an information processing device running an appropriate program, which will not be described in detail here.

[0175] The network environment awareness module 301 detects and outputs network environment characteristics, including spatial and temporal traffic features, based on the access request and the target network's response. The masquerade / spoofing module 302 generates a masquerade strategy template for masquerading as normal network access and a spoofing strategy template for deceiving the target network, based on the network environment characteristics. The proxy pool optimization module 303 generates network access traffic and sends it to the target network based on the masquerade and spoofing strategy templates.

[0176] Corresponding to the second embodiment described above, as an example, such as Figure 4 The proxy IP optimization device shown may also include a risk energy detection module 304, which detects the risk energy or risk energy change rate that indicates the failure risk of the proxy IP. When the risk energy detection module 304 detects that the risk energy or risk energy change rate exceeds a preset threshold, the spoofing module 302 changes the strategy exploration level and template perturbation amplitude.

[0177] Specifically, the network environment perception module 301 can be configured to perform the processing corresponding to the network environment perception step S1 in the aforementioned proxy IP optimization method, the masquerade inducement module 302 can be configured to perform the processing corresponding to the masquerade inducement step S2 in the aforementioned proxy IP optimization method, the proxy pool optimization module 303 can be configured to perform the processing corresponding to the proxy pool optimization step S3 in the aforementioned proxy IP optimization method, and the risk energy detection module 304 can be configured to perform the processing corresponding to the risk energy detection step S4 in the aforementioned proxy IP optimization method. These details will not be repeated here.

[0178] On the other hand, the present invention can also be implemented as a computer device. Figure 5 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention, such as... Figure 5 As shown, the computer device 500 includes a memory 510, a processor 520, and a bus 530. The processor 520 and the memory 510 communicate with each other via the bus 530. The memory 510 stores a computer program 511, and the processor 520 calls the computer program 511 stored in the memory 510. The computer program 511 is programmed to implement the various steps of the proxy IP optimization method described above when executed by the processor 520.

[0179] The present invention can also be implemented as a computer-readable storage medium storing a computer program that, when executed by a processor, implements the various steps of the proxy IP optimization method described above.

[0180] The present invention can also be implemented as a computer program product, which includes a computer program that, when executed by a processor, implements the various steps of the proxy IP optimization method described above.

[0181] Those skilled in the art will understand from the foregoing description that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0182] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0183] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0184] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0185] The information collected in the technical solution of this application is information and data authorized by the user or fully authorized by all parties. The collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with the relevant laws, regulations and standards of the relevant countries and regions, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation portals are provided for users to choose to authorize or refuse.

[0186] The acquisition, transmission, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0187] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0188] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing proxy IPs, characterized in that... The following steps are included in the iterative execution: The network environment awareness step detects and outputs network environment features that include spatial and temporal traffic characteristics based on the access request and the target network's response. The masquerade and inducement step involves generating a masquerade strategy template for masquerading as normal network access and an inducement strategy template for inducing access to the target network based on the network environment characteristics; and The proxy pool optimization step involves generating network access traffic and sending it to the target network based on the masquerade policy template and the inducement policy template.

2. The proxy IP optimization method according to claim 1, characterized in that... It also includes the following steps: The risk energy detection step detects the risk energy or the rate of change of risk energy that indicates the failure risk of the proxy IP. as well as In the camouflage-inducing adjustment step, if the risk energy or the rate of change of risk energy exceeds a preset threshold, the strategy exploration level and template perturbation amplitude when generating the camouflage strategy template and the inducement strategy template in the camouflage-inducing step are adjusted.

3. The proxy IP optimization method according to claim 2, characterized in that, If the risk energy or the rate of change of the risk energy exceeds the preset threshold, the strategy exploration level is increased, and the perturbation amplitude of the camouflage strategy template and / or the inducement strategy template is increased.

4. The proxy IP optimization method according to any one of claims 1-3, characterized in that... It also includes the following steps: In the network environment perception step, a convolutional neural network is used to extract the traffic features in the spatial dimension, and a long short-term memory network is used to extract the traffic features in the temporal dimension.

5. The proxy IP optimization method according to any one of claims 1-3, characterized in that, In the proxy pool optimization step, access requests that mimic the traffic characteristics of normal users are generated based on the masquerade strategy template.

6. The proxy IP optimization method according to any one of claims 1-3, characterized in that, In the proxy pool optimization step, access requests for interfering with the detection model of the target website are generated based on the inducement strategy template.

7. The proxy IP optimization method according to any one of claims 1-3, characterized in that, In the network environment perception step, a network environment feature vector containing traffic features in the spatial dimension and traffic features in the temporal dimension is output. In the spoofing induction step, a generative adversarial network is used, with the network environment feature vector and the features representing the network response of the object as inputs to the generative adversarial network, and the induction strategy template is generated based on the output of the generative network in the generative adversarial network.

8. A proxy IP optimization device, characterized in that... include: The network environment awareness module detects and outputs network environment features that include spatial and temporal traffic characteristics based on access requests and the response from the target network. The masquerade and inducement module generates, based on the network environment characteristics, a masquerade strategy template for masquerading as normal network access and an inducement strategy template for inducing access to the target network; and The proxy pool optimization module generates network access traffic and sends it to the target network based on the masquerade policy template and the inducement policy template.

9. The proxy IP optimization device according to claim 8, characterized in that... Also includes: The risk energy detection module detects the risk energy or the rate of change of risk energy, which indicates the failure risk of the proxy IP. When the risk energy detection module detects that the risk energy or the rate of change of the risk energy exceeds a preset threshold, the camouflage induction module changes the strategy exploration level and template perturbation amplitude when generating the camouflage strategy template and the induction strategy template.

10. The proxy IP optimization device according to claim 9, characterized in that, When the risk energy detection module detects that the risk energy or the rate of change of the risk energy exceeds the preset threshold, the camouflage-inducing module increases the degree of strategy exploration and increases the perturbation amplitude of the camouflage strategy template and / or the inducing strategy template.

11. The proxy IP optimization apparatus according to any one of claims 8-10, characterized in that, The network environment perception module uses a convolutional neural network to extract the traffic features in the spatial dimension and a long short-term memory network to extract the traffic features in the temporal dimension.

12. The proxy IP optimization device according to any one of claims 8-10, characterized in that, The proxy pool optimization module generates access requests that mimic the traffic characteristics of normal users based on the masquerade strategy template.

13. The proxy IP optimization device according to any one of claims 8-10, characterized in that, The proxy pool optimization module generates access requests based on the inducement strategy template to interfere with the detection model of the target website.

14. The proxy IP optimization device according to any one of claims 8-10, characterized in that, The network environment perception module outputs a network environment feature vector that includes traffic features in the spatial dimension and traffic features in the temporal dimension. The camouflage inducement module utilizes a generative adversarial network (GAN) to take the network environment feature vector and the features representing the network response of the target as inputs to the GAN, and generates the inducement strategy template based on the output of the generative network in the GAN.

15. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the proxy IP optimization method according to any one of claims 1 to 7.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the proxy IP optimization method according to any one of claims 1 to 7.

17. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the proxy IP optimization method according to any one of claims 1 to 7.

Images