A Machine Learning-Based Blockchain Transaction Obfuscation Method

Abstract

This invention relates to a machine learning-based blockchain transaction obfuscation method, belonging to the field of blockchain data covert transmission technology. This method, based on generative adversarial networks (GANs), trains covert transaction fields that are difficult for detection models to distinguish, constructing covert transactions that are difficult for machine learning detection models to detect. This invention is the first to propose using machine learning methods for covert transaction construction, which better simulates the field characteristics of real blockchain transactions during the construction process, significantly improving the covertness of the generated transactions. Simultaneously, machine learning methods can generate transactions in large batches, improving the efficiency of covert transaction construction. Transactions generated using GANs will be difficult to identify by the same neural network model subsequently, achieving the goal of obfuscating real transactions from covert transactions.

Description

Technical Field

[0001] This invention relates to a data covert transmission obfuscation method, specifically a blockchain transaction obfuscation method based on machine learning, belonging to the field of blockchain and data information covert transmission technology. Background Technology

[0002] With information security issues becoming increasingly prominent, data encryption and covert data transmission technologies have become the main technical means to effectively solve the problem of information theft. Unlike encryption technology, which focuses on the information itself and its readability, covert data transmission technology aims to hide communication behavior and the identities of the communicating parties. It hides the information that needs to be transmitted confidentially within a normal digital carrier. These carriers containing information are often very similar to the original carrier, making it difficult for third parties to detect the existence of secret information, thus enabling secure data transmission in a comprehensive and intensive monitoring environment.

[0003] Covert data transmission technology involves the sender, receiver, and communication channel of the message. This technology holds a crucial position in the field of information security, especially given the increasingly fierce nature of current network attack and defense tactics; therefore, research into covert data transmission technology, or covert communication, is of great significance.

[0004] While traditional covert channel schemes can transmit information covertly, they still have obvious characteristics. Attackers can use certain techniques (such as traffic analysis attacks and eavesdropping) to discover communication traffic, associate communication links, and then associate the real identities of the communicating parties to restore or even tamper with the communication content, making it difficult to achieve the requirements of covert communication.

[0005] In recent years, the rapid rise of decentralized blockchain technology has brought new possibilities to data concealment technology. Combining traditional data encryption and information hiding techniques with the new blockchain transmission channel is an important direction for future research. Blockchain technology possesses the characteristics of decentralization, immutability, unforgeability, openness, and security. Each node on the blockchain has a backup copy of the on-chain data, and every user can access all data on the chain, ensuring transparency. Blockchain uses cryptographic technologies such as hash functions and digital signatures to guarantee the immutability of on-chain data, improving data trustworthiness.

[0006] Based on these characteristics, blockchain offers a superior alternative to covert communication. It implements a public, distributed ledger built on a peer-to-peer network, which can be viewed as a bulletin board simulating a natural broadcast channel. In this communication model, message transmission has no specific destination, perfectly concealing the recipient's identity. The sender simply generates an address locally as an account and packages the data into a transaction. This anonymity helps hide the true identities of both communicating parties. Furthermore, the immutability of blockchain makes it difficult for attackers to disrupt the transmission of secret messages.

[0007] While using blockchain and cryptographic encryption technologies for covert information transmission can greatly improve the confidentiality of information, problems such as information leakage still exist. The main reasons are as follows:

[0008] 1. The covert transmission of information using blockchain has only gradually emerged in recent years. Currently, there are very few covert communication solutions based on blockchain, and some solutions are relatively simple and easy to detect.

[0009] 2. When constructing covert transactions on the blockchain, simply embedding covert information into existing blockchain transactions or simulating a small number of existing transactions to construct covert transactions is insufficient to create covert transactions with the same characteristics as real blockchain transactions. This leads to the identification of covert transactions with embedded information and the leakage of covert information. Summary of the Invention

[0010] The purpose of this invention is to address the challenges posed by deep learning-based detection models, improve the concealment of blockchain data concealed transactions, and creatively propose a machine learning-based blockchain transaction obfuscation method.

[0011] This method is based on generative adversarial networks, which train hidden transaction fields that are difficult for detection models to distinguish, thus constructing hidden transactions that are difficult for machine learning detection models to detect.

[0012] First, the technical terms involved in this invention will be explained and described.

[0013] Blockchain network: refers to a blockchain network used for transmitting data. There are three types of blockchain networks: public blockchain, consortium blockchain, and private blockchain.

[0014] Blockchain transaction: A blockchain transaction refers to a transaction used to store data. A blockchain transaction is a data structure used by different nodes in a blockchain network to complete data interaction. Each blockchain transaction contains a sender address field, a receiver address field, a data storage field, and other fields. Blockchain transactions are transmitted to all nodes in the network in a flooding broadcast manner.

[0015] Covert transactions refer to special blockchain transactions that embed covert information. These transactions are generated by the covert communicator and published to the blockchain network. The covert communicator's receiver identifies the covert transaction from among numerous transactions on the blockchain network using a specific filtering method, and extracts the covert information carried in the transaction.

[0016] Generative Adversarial Networks (GANs) are networks proposed by Goodfellow et al. A GAN consists of two neural networks: a generator (G) and a discriminator (D). It learns the distribution of real training data through a game between the generator and the discriminator, generating simulated data that resembles real data.

[0017] The objective of this invention is achieved through the following technical solution.

[0018] A machine learning-based blockchain transaction obfuscation method includes the following steps:

[0019] Step 1: Based on the hidden transaction type and characteristics to be embedded, filter and download ordinary transactions (e.g., 1000) on a public blockchain network.

[0020] Step 2: Extract the unembedded fields from ordinary transactions and construct an input sequence that meets the model requirements.

[0021] Step 3: Perform word embedding processing on the input sequence.

[0022] Step 4: Put the word-embedded sequence into the generator inside the model, and the generator will generate a simulated sequence.

[0023] Step 5: Use the model's internal discriminator to evaluate the simulated sequence and the input sequence, and generate feedback to the generator.

[0024] Step 6: The generator receives feedback, updates parameters, and generates a new simulated sequence.

[0025] Step 7: Continuously iterate the generator and discriminator, and finally generate an output sequence that the adversarial model cannot distinguish from the input sequence.

[0026] Step 8: Construct covert transactions by combining the output sequence generated by the model with the covert communication construction scheme.

[0027] Beneficial effects

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] 1. Enhanced Concealment. This invention is the first to propose using machine learning methods for constructing covert transactions. During the construction process, it better simulates the field characteristics of real blockchain transactions, significantly improving the concealment of the generated transactions. Furthermore, the use of machine learning methods allows for large-scale transaction generation, increasing the efficiency of constructing covert transactions.

[0030] 2. Enhanced resistance to detection. The process of generating transaction fields using generative adversarial networks is a continuous process of detection and updating. Transactions generated using generative adversarial networks are difficult to identify by the same neural network model, thus achieving the goal of confusing real transactions with hidden transactions. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall structure of the Covert Transaction Model (PCTC) of this invention.

[0032] Figure 2 This is a schematic diagram of the LSTM structure. Detailed Implementation

[0033] The method of the present invention will now be described more clearly and completely with reference to the accompanying drawings.

[0034] A machine learning-based blockchain transaction obfuscation method includes the following steps:

[0035] Step 1: Based on the type and characteristics of the hidden transactions into which information is to be embedded, filter and download ordinary transactions on a public blockchain network.

[0036] Step 2: Extract the unembedded fields from ordinary transactions and construct an input sequence that meets the model requirements.

[0037] Step 3: Perform word embedding processing on the input sequence.

[0038] Step 4: Put the word-embedded sequence into the generator inside the model, and the generator will generate a simulated sequence.

[0039] Step 5: Use the model's internal discriminator to evaluate the simulated sequence and the input sequence, and generate feedback to the generator.

[0040] Step 6: The generator receives feedback, updates parameters, and generates a new simulated sequence.

[0041] Step 7: Continuously iterate the generator and discriminator, and finally generate an output sequence that the adversarial model cannot distinguish from the input sequence.

[0042] Step 8: Construct covert transactions by combining the output sequence generated by the model with a specific covert communication construction scheme.

[0043] like Figure 1 As shown, this invention proposes a practical covert transaction construction model, abbreviated as PCTC. This model is based on a generative adversarial network (SeqGAN) for sequence generation, including a generator and a discriminator.

[0044] The generator generates certain hidden transaction fields used for generating hidden transactions.

[0045] Specifically, the generator can be implemented using a Long Short-Term Memory (LSTM) network. The discriminator is used to determine whether the input sequence is a field from a normal blockchain transaction or a field generated by the generator.

[0046] The PCTC training process includes:

[0047] A sentence start marker (BOS) is input into the generator network. In this model, this marker is "BOS". After iteration through the LSTM network, the fields required for the covert transaction are generated, and these fields are then input into the discriminator. The discriminator extracts features from the fields and determines whether the input is a field from a normal transaction or a field generated by the generator.

[0048] When the generator was initially trained, its generation performance was poor, with significant differences between the generated fields and those of normal transactions. In contrast, the discriminator easily distinguished between the generated and real fields, achieving high accuracy. Therefore, the discriminator's results were fed back to the generator, guiding it to adjust its weights and continuously improve its generation performance, making the generated fields increasingly difficult to differentiate from normal fields.

[0049] As the generator and discriminator continuously compete with each other, their respective effects will gradually improve. The fields generated by the generator will be closer to the real fields, and the discriminator's discrimination ability will also be continuously improved.

[0050] Specifically, the technical principles of the generator and discriminator, as well as the overall implementation of the model, are as follows:

[0051] Generator: Long Short-Term Memory (LSTM) networks can be used as generator networks.

[0052] Long Short-Term Memory (LSTM) networks are a variant of Recurrent Neural Networks (RNNs). They introduce gating mechanisms to control the rate of information accumulation, selectively adding or forgetting information. They can also be called gated recurrent neural networks, aiming to improve the long-range dependency problem of simple recurrent neural networks. RNNs are a type of neural network used to process temporal data. Words in sequential data are often not independent; predicting the next word usually requires considering not only the current word but also earlier words. Therefore, RNNs, by using self-feedback neurons, leverage their ability to extract deep temporal and semantic information from the data, enabling them to process temporal data of arbitrary length. While RNNs can theoretically establish dependencies between states over long time intervals, they suffer from vanishing or exploding ladder problems during learning, practically only learning short-term dependencies.

[0053] To address this problem inherent in recurrent neural networks (RNNs), gated Long Short-Term Memory (LSTM) networks were proposed. LSTMs outperform RNNs in handling long sequences. The working mechanism and computational process of LSTMs are as follows: Figure 2 As shown in Figure 1, and Equations 1-6, the LSTM network introduces a new internal state. This is used for linear recurrent information transfer, which includes information remembered by the network, while nonlinear output information is given to the external state of the hidden layer. .

[0054] LSTM introduces a gating mechanism to control the information transmission path, which uses three gates: the input gate, the input gate, and the input gate. Forgotten Gate and output gate Current input and the external state at the previous moment The input data for each loop unit is processed through three gates.

[0055] First, pass through the Gate of Oblivion. , Determine the internal state at the previous moment Which information in the data should be forgotten? Output the filtered information. As shown in Equation 1.

[0056] Next is the input gate. Input gate Determine the candidate state at the current moment The amount of information that needs to be saved is shown in Equation 2.

[0057] As shown in Equation 4, the candidate state It is obtained by applying the tanh activation function to the input data. The internal state at the previous time step. Output of the forget gate Candidate states for element-wise multiplication Input gate Element-wise multiplication, adding the two results, yields the current internal state. As shown in equation 5.

[0058] Finally, it passes through the output gate. Output gate Determine the internal state at the current moment How much information needs to be output to the external state? As shown in Equation 3.

[0059] As shown in Equation 6, the internal state at the current time... After the tanh activation function and the output of the output gate Multiplying these yields the external state at the current moment. The "gate" in the LSTM network uses σ(·), which is the Logistic function, taking values ​​between (0,1), representing the percentage of information allowed to pass through.

[0060] Given an input sequence LSTM evaluates the input-to-predict-output sequence by iterating through the following equations. Mapping:

[0061] (1)

[0062] (2)

[0063] (3)

[0064] (4)

[0065] (5)

[0066] (6)

[0067] (7)

[0068] in, , , These are the weights for forgetting, input, and output gates to input, respectively. , , These are the weight matrices from the forgotten input, output gate to the hidden state. , , It is a bias vector related to forgetting, input, and output gates. , This represents the weight moments of the tanh activation function. This represents the bias vector of the tanh activation function. The predicted state from the LSTM is calculated using Equation 7. and This represents the output weight matrix and bias vector.

[0069] Discriminator: An LSTM network is used as the discriminator network.

[0070] To improve performance, this method further incorporates a Highway Network based on pooling feature maps. The Highway Network simplifies gradient-based training of deep neural networks, allowing information to flow rapidly and unimpeded through each layer, effectively mitigating gradient problems and enabling deep neural networks to perform beyond the limitations of shallow neural networks. Inspired by LSTM, this network adds gate functions T (transform gate) and C (carry gate), where T represents the transformed portion of the input information and C represents the retained portion of the original input information. Due to this gating mechanism, the neural network has a path, allowing information to flow along multiple layers without attenuation. As shown in Equation 8, the external state from LSTM... It will be further processed by the Highway Network and the final output will be... It consists of two parts: one part is the predicted state transformed from the input information, and the other part is the input information itself. and These are the weights corresponding to the gate functions T and C, respectively.

[0071] For simplicity, let's set C = 1 – T.

[0072] (8)

[0073] in, This indicates the final output. Indicates the external state. This represents the input at the current moment.

[0074] Finally, a fully connected layer is added, using the sigmoid activation function to output the probability that the input sequence is true.

[0075] To enable the generator to produce more realistic transaction fields, a hidden transaction detection model consisting of a Text Convolutional Neural Network (TextCNN) and a Backpropagation Neural Network (BPNN) can be incorporated into the discriminator. This detection model is trained using known hidden transactions. The hidden transaction detection model and the LSTM model jointly calculate the reward, which is then fed back to the generator. During training, the weights of the LSTM model change, but the weights of the hidden transaction detection model remain constant.

[0076] Model implementation and loss function:

[0077] like Figure 1 As shown, the input to the generator network is a sentence start marker "BOS". This marker is converted into a word vector and fed into the generator LSTM network. The output of the generator at time t is... As input for the next moment In this way, the target sequence is continuously generated from a starting marker. Where γ is a vocabulary containing all possible generated words.

[0078] The sequence generated at time t Let it be denoted as state s, the next candidate word. This is denoted as action 'a'. Each time the generator generates a new word... All of them will generate the sequence The reward is evaluated within the discriminator network.

[0079] The discriminator can directly evaluate the reward for a complete sentence. However, to evaluate the reward for an intermediate state sequence (a partial sequence), the model uses Monte Carlo search to randomly generate Tt unknown words, completes the sentence, and then feeds it into the discriminator to evaluate the reward. The reward from this process is accumulated, and after the generator completes the entire sentence, the weights are updated using the reward given by the discriminator. Specifically, the generator is considered to start from the initial state... The model for generating sequences uses a strategy that maximizes Equation 9:

[0080] (9)

[0081] in, Represents the cost function; Represents the mathematical expectation; The reward is for the complete sequence, and the reward comes from the discriminator. The discriminator consists of real data and a generator. The generated data are used together for training; and These are the parameters of the discriminator and the generator; It is the action-value function of the sequence, defined as:

[0082]

[0083] (10)

[0084] in, It equals the total sequence length minus 1. Set the discriminator. Half of the results come from the LSTM discriminant network, and the other half from the covert transaction detection model. The discriminant's loss function... It can be expressed as the following formula:

[0085] (11)

[0086] in, This represents a covert transaction detection model;

[0087] The probability that the discriminator expects to be the true sample As large as possible. Therefore, minimize the loss function. Then, the loss function of the PCTC model is expressed as follows:

[0088] (12)

[0089] The discriminator D desires the probability of the true sample to be as high as possible, and the probability of the sample generated by the generator to be as low as possible. Equation 12 aims to maximize the probability of the true sample for the discriminator D. Conversely, the generator G desires the probability of the generated sample to be as high as possible, and Equation 12 aims to minimize the probability of the generated sample for the generator G.

Claims

1. A blockchain transaction obfuscation method based on machine learning, characterized in that, Includes the following steps: Step 1: Based on the type and characteristics of the hidden transactions into which information is to be embedded, filter and download ordinary transactions on a public blockchain network; Step 2: Extract the non-embedded fields from ordinary transactions to construct an input sequence that meets the model requirements; The covert transaction construction model is based on a generative adversarial network for sequence generation, which includes a generator and a discriminator. The implementation of the covert transaction model and its loss function are as follows: The generator network takes a sentence start marker "BOS" as input, converts this marker into word vectors, and feeds them into the generator network. The generator's output at time t is... As input for the next moment In this way, the target sequence is continuously generated from a starting marker. Where γ is a vocabulary containing all possible generated words; The sequence generated at time t Let this be denoted as state s, the next candidate word. This is denoted as action a; each time the generator generates a new word... All of them will generate the sequence The reward is evaluated within the discriminator network. The discriminator can directly evaluate the reward of the complete sentence. To evaluate the reward of the intermediate state sequence, the model uses Monte Carlo search to randomly generate Tt unknown words, completes the sentence, and puts it into the discriminator to evaluate the reward. The reward of this process is accumulated. After the generator completes the whole sentence, the reward given by the discriminator is used to update the weight. Treat the generator as starting from the initial state. A model for generating sequences, employing a strategy that maximizes the following: in, Represents the cost function; Represents the mathematical expectation; The reward is for the complete sequence, and the reward comes from the discriminator. The discriminator consists of real data and a generator. The generated data are used together for training; and These are the parameters of the discriminator and the generator; It is the action-value function of the sequence, defined as: in, Equals the total sequence length - 1; sets the discriminator. Half of the results come from the discriminant network, and half from the covert transaction detection model. The discriminant's loss function... It can be expressed as the following formula: in, This represents a covert transaction detection model; The probability that the discriminator expects to be the true sample As large as possible; therefore, minimize the loss function. Then, the loss function of the PCTC model is expressed as follows: Discriminator D wants the probability of the real sample to be as high as possible and the probability of the sample generated by the generator to be as low as possible. The above formula is to maximize the probability of the real sample for discriminator D. Conversely, generator G wants the probability of the generated sample to be as high as possible. The above formula is to minimize the probability of the generated sample for generator G. Step 3: Perform word embedding processing on the input sequence; Step 4: Feed the word-embedded sequence into the generator inside the model, and the generator will generate a simulated sequence; Step 5: Use the model's internal discriminator to evaluate the simulated sequence and the input sequence, and generate feedback to the generator; Step 6: The generator receives feedback, updates parameters, and generates a new simulated sequence; Step 7: Continuously iterate through the generator and discriminator to finally generate an output sequence that the adversarial model cannot distinguish from the input sequence; Step 8: Construct covert transactions by combining the output sequence generated by the model with the covert communication construction scheme.

2. The blockchain transaction obfuscation method based on machine learning as described in claim 1, characterized in that, in, The generator generates partial covert transaction fields used for covert transaction generation; The discriminator is used to determine whether the input sequence is a field in a normal blockchain transaction or a field generated by the generator; The process of training a covert transaction model in one round includes: A sentence start marker is input into the generator network. In this model, the marker is "BOS". After iteration through the LSTM network, the field required for the covert transaction is generated and input into the discriminator. The discriminator extracts features from the field and determines whether the input is a field from a normal transaction or a field generated by the generator. The discriminator's results are fed back to the generator, guiding the generator to adjust its weights.

3. The blockchain transaction obfuscation method based on machine learning as described in claim 2, characterized in that, The Long Short-Term Memory (LSTM) network is used as the generator network.

4. The blockchain transaction obfuscation method based on machine learning as described in claim 2, characterized in that, An LSTM network is used as the discriminator network, and a Highway Network based on pooling feature maps is added. The Highway Network simplifies gradient-based training of deep neural networks, allowing information to pass through the layers of deep neural networks at high speed without obstruction. Inspired by LSTM, the Highway Network adds gate functions T and C, where T represents the part of the input information that is transformed and C represents the part of the original input information that is retained. External state from LSTM It will be further processed by the Highway Network, and the final output will be... It consists of two parts: one part is the predicted state transformed from the input information, and the other part is the input information itself. and These are the weights corresponding to the gate functions T and C, respectively; Set C = 1 – T; in, This indicates the final output. Indicates the external state. This represents the input at the current moment; Finally, a fully connected layer is added, using the sigmoid activation function to output the probability that the input sequence is true.

5. The blockchain transaction obfuscation method based on machine learning as described in claim 4, characterized in that, A covert transaction detection model consisting of a text convolutional neural network and a backpropagation neural network is added to the discriminator. This detection model is trained using known covert transactions. The covert transaction detection model and the LSTM model jointly calculate the reward, which is used to feed back to the generator. During training, the weights of the LSTM model will change, but the weights of the covert transaction detection model will remain fixed.

Images