Blockchain gambling behavior identification method and system based on cross-modal retrieval

Abstract

The application discloses a blockchain gambling behavior identification method and system based on cross-modal retrieval and belongs to the field of blockchain security. The method specifically comprises the following steps: (1) smart contract gambling behavior identification: using smart contract bytecode to identify smart contracts related to gambling; (2) address gambling behavior identification: using graph information extraction technology to identify addresses related to gambling contracts; (3) gambling behavior identification result correction: according to the address gambling behavior identification result, using a feedback correction method to correct the smart contract gambling behavior identification result obtained in step (1). The method provides a novel gambling behavior identification scheme for blockchain smart contracts and addresses, and can accurately identify gambling behavior in a blockchain network.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of blockchain security, and particularly relates to a blockchain gambling behavior identification method and system based on cross-modal retrieval. BACKGROUND

[0002] In recent years, the blockchain technology has developed rapidly and has attracted widespread attention from all walks of life. The blockchain is essentially a distributed ledger that is shared among all participating nodes in a peer-to-peer network. The most critical feature of the blockchain is that it can maintain a secure and immutable transaction ledger in a peer-to-peer network that is mutually untrusted. The emergence of the blockchain technology provides an execution environment for smart contracts. A smart contract is a program stored on the blockchain, which can automatically execute according to pre-set conditions, and the execution is deterministic and irreversible, thereby ensuring the fairness of all participants.

[0003] Traditional gambling websites have problems such as opaque gambling process, fictitious prize pool, and refusal to pay winners, etc. In contrast, smart contract-based gambling applications strictly execute gambling rules according to pre-defined rules, and the entire process is completely transparent. Therefore, the number of smart contract-based gambling applications has grown rapidly in the past few years, which further puts strict requirements on the identification and regulation of gambling behaviors on the blockchain.

[0004] At present, the number of entities and transactions in the blockchain network has reached hundreds of millions, and is still rising sharply. By analyzing these entities and transactions, users can clearly perceive the gambling risks existing in the blockchain, thereby protecting their own rights and interests; in addition, identifying gambling behaviors of entities and transactions in the blockchain can also provide a novel security protection scheme for the entire blockchain ecosystem. However, there is currently no technical solution that can automatically identify gambling behaviors on the blockchain, mainly due to the following difficulties:

[0005] (1) The data of transactions, accounts, smart contracts, etc. in the blockchain are heterogeneous and multi-modal, and to analyze gambling behaviors, it is necessary to combine multiple entity types such as transactions, address accounts, and smart contracts. At present, there is still a lack of targeted cross-modal data extraction methods.

[0006] (2) Smart contracts are compiled into bytecode before being deployed to Ethereum. Bytecode in assembly language form is different from source code, and it is composed of a series of binary codes, which is difficult to understand and analyze directly using tools. Existing methods usually analyze bytecode using manual experience and decompilation, which on the one hand cannot accurately capture the semantic information of smart contracts, and on the other hand will produce a large number of false positives.

[0007] (3) The transaction data of different addresses has a large size difference. The number of transactions involving an address usually ranges from one to tens of thousands. To analyze the transaction data, a general feature extraction framework needs to be designed for addresses with a large transaction size difference. SUMMARY

[0008] In view of the difficulties in the identification technology for large-scale gambling behavior in a blockchain, the application provides a blockchain gambling behavior identification method and system based on cross-modal retrieval. The application collects large-scale heterogeneous data in a blockchain network, introduces a decompilation technology and a graph information extraction technology to process the data, wherein the decompilation technology is used to extract features of a smart contract, and the graph information extraction technology is used to extract transaction features of an address; finally, a cross-modal retrieval technology is used to identify a smart contract and an address involved in gambling.

[0009] The technical solution adopted by the application is as follows:

[0010] The first object, the application provides a blockchain gambling behavior identification method based on cross-modal retrieval, comprising the following steps:

[0011] (1) Smart contract gambling behavior identification: using smart contract bytecode to identify smart contracts related to gambling;

[0012] (2) Address gambling behavior identification: using a graph information extraction technology to identify addresses related to gambling contracts;

[0013] (3) Gambling behavior identification result correction: according to the address gambling behavior identification result, using a feedback correction method to correct the smart contract gambling behavior identification result obtained in step (1).

[0014] Further, the step (1) comprises:

[0015] (1-1) decompiling the smart contract bytecode through a decompilation technology to obtain a smart contract operation code sequence;

[0016] (1-2) extracting operation codes related to gambling behavior from the smart contract operation code sequence, and constructing a feature vector of the smart contract according to the number of occurrences of each operation code related to gambling behavior;

[0017] (1-3) using a LightGBM classification model based on a memory component to classify the feature vector of the smart contract, and the classification types include gambling contracts and non-gambling contracts.

[0018] Further, the step (2) comprises:

[0019] (2-1) Collecting addresses that have traded with gambling contracts and corresponding transaction amounts, and constructing a transaction graph for each transaction address;

[0020] (2-2) Extracting transaction feature vectors from the transaction graph, including basic index features, degree index features, and amount index features; the basic index features are composed of the number of edges, the number of points, the number of input edges, and the number of output edges of the transaction graph; the degree index features are composed of the degree, the in-degree, and the out-degree of the transaction graph; the amount index features are composed of the total amount, the total input amount, the total output amount, the average amount, the average input amount, the average output amount, the input amount variance, and the output amount variance;

[0021] (2-3) Classifying the transaction feature vectors of the transaction graph using a LightGBM classification model based on a memory component, and the classification types include gambling addresses and non-gambling addresses.

[0022] Further, the construction method of the transaction graph is: converting each transaction address into a point, if the transaction address is a payment address in a transaction, constructing an output edge starting from the transaction address; on the contrary, if the transaction address is a receiving address, constructing an input edge pointing to the transaction address; the weight of the edge is the transaction amount.

[0023] Further, the training process of the LightGBM classification model based on the memory component is:

[0024] a. In the first round of training, use all training samples to train the LightGBM classification model;

[0025] b. Select the misclassified samples and their corresponding labels into the memory component;

[0026] c. In the next round of training, first train the LightGBM classification model for k epochs using all training samples, and then add an additional epoch to replay the misclassified samples recorded in the memory component, the construction method of the additional epoch is: assuming that the number of training samples in each epoch is n, randomly select n samples from all misclassified samples in the memory component to construct the additional epoch, if the number of misclassified samples is less than n, randomly select the remaining from all training samples;

[0027] d. Repeat steps b to c until the iteration is complete.

[0028] Further, the feedback correction method in step (3) is: if the proportion of addresses identified as gambling addresses in all addresses related to gambling contracts does not exceed a threshold value, the identification result of the gambling contract is corrected to a non-gambling contract.

[0029] The second objective of this invention is to provide a blockchain gambling behavior identification system based on cross-modal retrieval technology, comprising:

[0030] The smart contract gambling behavior identification module is used to identify smart contracts related to gambling by utilizing the bytecode of smart contracts;

[0031] The address gambling behavior identification module is used to identify addresses related to gambling contracts using graph information extraction technology;

[0032] The gambling behavior identification result correction module is used to correct the identification results of the smart contract gambling behavior identification module based on the identification results of the address gambling behavior identification module using a feedback correction method.

[0033] As a preferred embodiment of the present invention, the smart contract gambling behavior identification module includes:

[0034] The EVM decompiler is used to decompile smart contract bytecode using decompilation technology to obtain the smart contract opcode sequence.

[0035] An operator extractor is used to extract opcodes related to gambling behavior from the opcode sequence of a smart contract and construct a feature vector of the smart contract based on the frequency of occurrence of each opcode related to gambling behavior.

[0036] A smart contract classifier that uses a memory-based LightGBM classification model to classify the feature vectors of smart contracts, including gambling contracts and non-gambling contracts.

[0037] As a preferred embodiment of the present invention, the address gambling behavior identification module includes:

[0038] A transaction graph builder, which collects addresses that have transacted with gambling contracts and the corresponding transaction amounts, and constructs a transaction graph for each transaction address;

[0039] A graph structure extractor is used to extract transaction feature vectors from a transaction graph, including basic indicator features, degree indicator features, and monetary indicator features. The basic indicator features consist of the number of edges, the number of vertices, the number of input edges, and the number of output edges in the transaction graph. The degree indicator features consist of the degree, in-degree, and out-degree of the transaction graph. The monetary indicator features consist of the total monetary amount, the total input monetary amount, the total output monetary amount, the average monetary amount, the average input monetary amount, the average output monetary amount, the variance of the input monetary amount, and the variance of the output monetary amount.

[0040] An address behavior classifier is configured to classify transaction feature vectors of a transaction graph by using a LightGBM classification model based on a memory component, and the classification types include gambling addresses and non-gambling addresses.

[0041] As a preferred embodiment of the present application, the feedback correction method in the gambling behavior identification result correction module is: if the proportion of addresses identified as gambling addresses in all addresses related to the gambling contract does not exceed a threshold value, the identification result of the gambling contract is corrected to a non-gambling contract.

[0042] The present application has the following contributions in the existing research on blockchain gambling entity behavior identification:

[0043] (1) The present application first proposes a new problem of detecting gambling contracts and gambling addresses from blockchain multi-modal data, and proposes the first accurate and efficient solution to the problem, specifically designs two feature extraction methods for multi-modal data, namely smart contract bytecode feature extraction paradigm and Ethereum address transaction feature extraction paradigm, and realizes high-precision identification of contracts and addresses participating in gambling. Through verification in a million-level data set, the F1-score index of the gambling address and gambling contract tasks proposed in the present application reaches 0.72 and 0.89 respectively, fully demonstrating the effectiveness of the present application.

[0044] (2) The present application innovatively proposes a cross-modal classification result correction mechanism. The correction mechanism can use the classification results of gambling addresses to correct the classification results of gambling contracts, and can improve the accuracy of classification without reducing the recall rate, thereby improving the overall identification accuracy of the framework. By using the classification result correction mechanism, the F1-score index of the gambling contract classification is increased from 0.82 to 0.89, fully demonstrating the effectiveness of the present application.

[0045] (3) The present application innovatively proposes a LightGBM classification model with a memory component. In the training process, the classification model selects the samples that are misclassified in the last round of training and uses them to construct the sample subset in the next round of training, which can further improve the classification accuracy of gambling behavior identification. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The overall design structure diagram of the present application.

[0047] Figure 2 The identification step diagram of the present application.

[0048] Figure 3 The transaction graph structure diagram of the present application. DETAILED DESCRIPTION

[0049] For the purpose of clearly explaining the present application, making the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application is described clearly and completely below in combination with the drawings in the embodiments of the present application, so that the person skilled in the art can implement according to the description. The drawings in combination with the specific implementation will be described in detail below.

[0050] The overall design structure diagram of the present application is shown in Figure 1 The focus is divided into two parts: gambling contract classification and gambling address classification. The gambling contract classification part includes an EVM disassembler, an operator extractor and a smart contract classifier; the gambling address classification part includes a transaction graph constructor, a graph structure extractor and an address behavior classifier.

[0051] First step: gambling contract classification

[0052] Component 1: EVM disassembler. Since most smart contracts exist in the form of bytecode, the present application implements an EVM disassembler to extract EVM opcodes from it.

[0053] Specifically, bytecode is an executable program composed of a series of <opcode, data> pairs. The Ethereum Virtual Machine (EVM) is the runtime environment for bytecode on Ethereum. In order to analyze the execution logic of the smart contract, the system needs to first parse the bytecode into binary code using the opcode rules of the EVM. The parsed binary code includes EVM opcodes and operands, and the EVM opcodes are divided into specific instruction sets, such as arithmetic operations, logical and comparison operations, control flow, system calls, stack operations and memory operations. In addition to typical bytecode operations, the EVM must also manage account information (e.g., address and balance), current gas price and block information. Table 1 lists several common Ethereum opcodes, each corresponding to a specific assembly language operation.

[0054] Table 1 Common Ethereum opcodes

[0055] Value mnemonic meaning Value mnemonic meaning 0x00 STOP stop execution 0x16 AND bitwise AND 0x01 ADD addition 0x17 OR bitwise OR 0x02 MUL multiplication 0x18 XOR bitwise XOR 0x03 SUB subtraction 0x19 NOT bitwise NOT 0x04 DIV integer division 0x20 SHA3 compute Keccak-256 hash 0x06 MOD Mod division remainder 0x30 ADDRESS address of the current executing account 0x10 LT less than 0x31 BALANCE balance of the executing account 0x11 GT greater than 0x32 ORIGIN from which account address the execution originated 0x14 EQ whether equal 0x33 CALLER caller address 0x15 ISZERO negation 0x35 CALLDATALOAD get the current environment input value

[0056] Component 2: Operator extractor. After disassembling the bytecode of the smart contract by the EVM disassembler, the system can obtain its corresponding EVM opcode sequence. Since the opcode sequence is unstructured data and the opcode sequences of different contracts have different lengths, the system needs to further encode the opcode sequence into a fixed-length feature vector.

[0057] The present application studies all 136 opcode operations in the EVM decompiler and calculates the occurrence of each opcode in each smart contract. Unlike other smart contracts, gambling contracts often involve random number generation operations and gambling fund collection operations. Through further observation, the present application finds that some opcodes rarely appear in contracts (e.g., PUSH5-PUSH32, DUP5-DUP16, SWAP5-SWAP16, etc.). Based on these observations, the system discards part of the rarely used and irrelevant to gambling opcodes. Finally, the system determines 80 opcode operations and uses the occurrence of each opcode operation in the smart contract as a feature.

[0058] Component 3: Smart contract classifier. After the smart contract binary program passes through the EVM disassembler and opcode feature extractor, the system can map each smart contract to a feature vector. Then, the system uses a classification model to distinguish the feature vectors of gambling contracts from those of other contracts. In this embodiment, the system selects LightGBM as the smart contract classifier of the system. At the same time, in order to make the model more intelligent, the present application innovatively proposes a LightGBM classification model based on a memory component.

[0059] The traditional LightGBM uses the negative gradient of the loss function as the residual approximation of the current decision tree, and then uses the residual approximation to fit a new decision tree. The working process of the LightGBM classification model based on the memory component proposed by the present application can be summarized as follows:

[0060] (1) In the first round of training, the system trains the LightGBM classification model using all training samples.

[0061] (2) The system selects all misclassified samples and their corresponding labels from the trained classification model and puts them into the memory component.

[0062] (3) In the next round of training, the system re-trains the classification model using all training samples. However, after every k epochs, the system adds an additional epoch to replay the misclassified samples recorded in the memory component. More specifically, assuming the number of training samples per epoch is n, the system randomly selects n samples from all misclassified samples in the memory component to construct the training subset for the next round, and if the number of misclassified samples is less than n, randomly selects the remaining from all training samples.

[0063] (4) Iterate steps (2) to (3) until convergence. By adding the memory component, the LightGBM model trained finally can improve the ability to learn from misclassified samples, and ultimately improve the accuracy of classification.

[0064] Second step: gambling address classification

[0065] Component 1: Transaction graph constructor. Unlike traditional transactions, the information available to the features of transactions on the blockchain platform is extremely limited. Therefore, the system only extracts the payment address, the receiving address and the transaction amount as the features of the transaction. For each address, the system obtains all addresses that have ever transacted with the address for modeling, and converts the transaction into an edge and the address information into a point to construct an address transaction graph. In this embodiment, if the address is a payment address in a transaction, an output edge is constructed from the address. Conversely, if the address is a receiving address, an input edge is constructed pointing to the address. The weight of the edge is set to the transaction amount. Figure 3 A typical transaction graph is shown.

[0066] Component 2: Graph structure extractor. In order to extract features from the transaction graph, the system uses three indicators, namely basic indicators, degree indicators and amount indicators, as shown in Table 2.

[0067] Table 2 Information extracted from the transaction graph

[0068]

[0069] Basic indicators: mainly focus on the overall structure of the graph. The indicators can capture the number of nodes, the number of edges and the edge type in the transaction graph.

[0070] Degree indicators: involve node information of the graph. The degree indicators mainly model the average degree, the average in-degree and the average out-degree of the nodes, where the out-degree of node i is the number of edges pointing from node i to other nodes, and the in-degree of node i is the number of edges pointing from other nodes to node i.

[0071] Amount indicators: focus on the information of the edges in the transaction graph. The amount indicators mainly include the total transaction amount of the edges, the average transaction amount and the transaction amount variance.

[0072] Component 3: Address behavior classifier. After obtaining the transaction graph features of each address, the system uses a machine learning model to classify the addresses, i.e. to classify them as gambling addresses and non-gambling addresses. In this embodiment, the same LightGBM model based on memory components as in the gambling contract classification part is used as the classifier, and the address data is used to train it.

[0073] Third step: correction of gambling behavior identification results

[0074] Due to the large number of contracts on Ethereum, gambling contracts only account for a small part of them. Therefore, even if the false positive rate of the smart contract classifier is very small, a large number of non-gambling contracts will be classified as gambling contracts. Therefore, the present invention designs a correction mechanism based on information feedback to further reduce the false positive rate of the smart contract classifier.

[0075] According to experience, if most of the addresses associated with the contract are not gambling addresses, the contract is most likely not a gambling contract. Through experimental comparison, the system finally selects a threshold of 80%, if less than 80% of the addresses associated with the contract are gambling addresses, the contract is most likely a false positive sample. After using this method, the false positive rate of the smart contract classifier is greatly reduced.

[0076] The above components constitute the blockchain gambling behavior identification system based on cross-modal retrieval of the present application. Embodiments of the system can be applied to any device with data processing capability, such as a computer or the like. The system embodiments can be implemented by software, or by hardware or a combination of software and hardware. For example, in one specific implementation of the present application, the blockchain gambling behavior identification method based on cross-modal retrieval includes the following steps, as shown in Figure 2

[0077] (1) Smart contract gambling behavior identification: identify smart contracts related to gambling behavior based on smart contract bytecode. Specifically, the bytecode is decompiled by decompilation technology to obtain smart contract operation code (Opcode); then, the operation code is defined with corresponding feature extraction rules, and a novel gambling behavior classifier is used, which is composed of a machine learning model and a memory component, and can identify gambling contracts.

[0078] (2) Blockchain address gambling behavior identification: after completing step (1), collect smart contracts involved in gambling, and find out addresses suspected of participating in gambling by capturing addresses that have traded with these smart contracts. In order to filter out the addresses that really participate in gambling from these addresses, a transaction graph is constructed for each address using graph information extraction technology, which simulates the transaction activities between the specified address and other addresses, thereby revealing the transaction characteristics of the address. At the same time, the transaction characteristics are classified using the gambling behavior classifier to identify gambling addresses.

[0079] (3) Gambling behavior identification result correction: a correction method based on information feedback is used to improve the overall accuracy of the gambling behavior identification system. The classification results of the addresses in step (2) are used to correct the classification results of the smart contracts in step (1), thereby improving the overall accuracy of the blockchain gambling behavior identification.

[0080] ​In the embodiment, it is verified in a data set of more than one million that the method reaches 0.72 and 0.89 in F1-score indexes of the two tasks of gambling address and gambling contract, fully illustrating the effectiveness of the method; if the correction method of information feedback is removed, the F1-score index of gambling contract classification is 0.82, and it can be seen that the F1-score index of gambling contract classification is increased from 0.82 to 0.89 by adopting the classification result correction mechanism, fully illustrating the effectiveness of the correction method of information feedback.

[0081] The above description of the embodiments is for facilitating those skilled in the art to understand and apply the present application. Those skilled in the art can easily make various modifications to the above embodiments, and apply the general principles described herein to other embodiments without creative labor. Therefore, the present application is not limited to the above embodiments, and the improvements and modifications of the present application made by those skilled in the art according to the disclosure of the present application should be within the protection scope of the present application.

Claims

1. A blockchain gambling behavior identification method based on cross-modal retrieval, characterized in that, The method comprises the following steps: (1) smart contract gambling behavior identification: using smart contract bytecode to identify smart contracts related to gambling; the step (1) comprises: (1-1) decompiling the smart contract bytecode through decompilation technology to obtain a smart contract operation code sequence; (1-2) extracting operation codes related to gambling behavior from the smart contract operation code sequence, and constructing a feature vector of the smart contract according to the number of occurrences of each operation code related to gambling behavior; (1-3) using a LightGBM classification model based on a memory component to classify the feature vector of the smart contract, and the classification types include gambling contracts and non-gambling contracts; (2) address gambling behavior identification: using graph information extraction technology to identify addresses related to gambling contracts; the step (2) comprises: (2-1) collecting addresses that have traded with gambling contracts and corresponding transaction amounts, and constructing a transaction graph for each transaction address; (2-2) extracting a transaction feature vector from the transaction graph, including basic index features, degree index features, and amount index features; the basic index features are composed of the number of edges, the number of points, the number of input edges, and the number of output edges of the transaction graph; the degree index features are composed of the degree, the in-degree, and the out-degree of the transaction graph; the amount index features are composed of the total amount, the total input amount, the total output amount, the average amount, the average input amount, the average output amount, the input amount variance, and the output amount variance; (2-3) using a LightGBM classification model based on a memory component to classify the transaction feature vector of the transaction graph, and the classification types include gambling addresses and non-gambling addresses; (3) gambling behavior identification result correction: according to the address gambling behavior identification result, using a feedback correction method to correct the smart contract gambling behavior identification result obtained in step (1).

2. The blockchain gambling behavior identification method based on cross-modal retrieval according to claim 1, characterized in that, The construction method of the transaction graph is: converting each transaction address into a point, if the transaction address is a payment address in a transaction, an output edge is constructed from the transaction address; on the contrary, if the transaction address is a receiving address, an input edge is constructed to the transaction address; the weight of the edge is the transaction amount. 3.The blockchain gambling behavior identification method based on cross-modal retrieval according to claim 1, characterized in that, The training process of the LightGBM classification model based on the memory component is: a. In the first round of training, use all training samples to train the LightGBM classification model; b. Select the misclassified samples and their corresponding labels into the memory component; c. In the next round of training, first train the LightGBM classification model for k epochs using all training samples, and then add an additional epoch to replay the misclassified samples recorded in the memory component, the construction method of the additional epoch is: assuming that the number of training samples in each epoch is n, randomly select n samples from all misclassified samples in the memory component to construct the additional epoch, if the number of misclassified samples is less than n, randomly select the remaining from all training samples; d. Repeat steps b to c until the iteration is completed.

4. The blockchain gambling behavior identification method based on cross-modal retrieval according to claim 1, characterized in that, The feedback correction method in step (3) is: if the proportion of addresses identified as gambling addresses in all addresses related to the gambling contract does not exceed a threshold value, the identification result of the gambling contract is corrected as a non-gambling contract. 5.A blockchain gambling behavior identification system based on cross-modal retrieval, characterized in that, The method comprises the steps of: an intelligent contract gambling behavior identification module for identifying gambling-related intelligent contracts using intelligent contract bytecode; an address gambling behavior identification module for identifying addresses related to gambling contracts using graph information extraction technology; a gambling behavior identification result correction module for correcting the identification result of the intelligent contract gambling behavior identification module according to the identification result of the address gambling behavior identification module using a feedback correction method; The intelligent contract gambling behavior identification module comprises: an EVM decompiler for decompiling the intelligent contract bytecode through a decompiling technology to obtain an intelligent contract opcode sequence; an operator extractor for extracting gambling behavior-related opcodes from the intelligent contract opcode sequence and constructing a feature vector of the intelligent contract according to the number of occurrences of each gambling behavior-related opcode; an intelligent contract classifier for classifying the feature vector of the intelligent contract using a LightGBM classification model based on a memory component, and the classification types include gambling contracts and non-gambling contracts; The address gambling behavior identification module comprises: a transaction graph constructor for collecting addresses that have transacted with the gambling contract and corresponding transaction amounts, and constructing a transaction graph for each transaction address; a graph structure extractor for extracting a transaction feature vector from the transaction graph, including basic index features, degree index features, and amount index features; the basic index features are composed of the number of edges, the number of points, the number of input edges, and the number of output edges of the transaction graph; the degree index features are composed of the degree, the in-degree, and the out-degree of the transaction graph; the amount index features are composed of the total amount, the total input amount, the total output amount, the average amount, the average input amount, the average output amount, the input amount variance, and the output amount variance; an address behavior classifier for classifying the transaction feature vector of the transaction graph using a LightGBM classification model based on a memory component, and the classification types include gambling addresses and non-gambling addresses. 6.The blockchain gambling behavior identification system based on cross-modal retrieval according to claim 5, characterized in that, The feedback correction method in the gambling behavior identification result correction module is: if the proportion of addresses identified as gambling addresses in all addresses related to the gambling contract does not exceed a threshold value, the identification result of the gambling contract is corrected as a non-gambling contract.

Images