Industrial big data privacy protection classification method based on federated learning

Abstract

This invention specifically relates to a privacy-preserving classification method for industrial big data based on federated learning, belonging to the field of industrial big data technology. It includes: a federated central server aggregating the ciphertext arrays of all active nodes element-by-element to generate a global ciphertext array; extracting the ciphertext elements of the hash index set corresponding to the industrial classification rules from the global ciphertext array and performing homomorphic aggregation; and introducing a blinding factor to generate blinded matching ciphertext. In this invention, based on the Paillier additive homomorphic encryption system, a full-process ciphertext computation system is constructed, encompassing local data feature encryption, global ciphertext aggregation, ciphertext rule matching, and blinded differential judgment. The entire process requires no decryption of the original data from any industrial node. Simultaneously, through a dual cryptographic blinding mechanism and threshold differential judgment logic, the decryption stage only outputs Boolean classification results, without disclosing any global statistical values ​​or local node data information, thus resisting various privacy theft risks such as differential attacks and data restoration attacks.

Description

Technical Field

[0001] This invention relates to the field of industrial big data technology, and in particular to a privacy protection classification method for industrial big data based on federated learning. Background Technology

[0002] The large-scale deployment of multi-source sensor networks in industrial sites has generated massive, multi-dimensional, and heterogeneous industrial time-series data. Collaborative analysis based on cross-node industrial big data has become a core technological support for achieving accurate equipment status classification, early fault warning, and production efficiency optimization.

[0003] The implementation of technology in industrial production sites faces multiple scenario-based constraints: Industrial edge nodes generally have limited computing power and storage resources. The industrial IoT environment also faces real-world problems such as bandwidth fluctuations, network latency, and occasional node offline events, which place extremely high demands on the lightweight, real-time performance, and robustness of analysis solutions. Industrial data is primarily composed of continuous physical quantities, and the requirements for classification accuracy and privacy protection vary across different scenarios. Solutions need to balance controllable accuracy, computational efficiency, and end-to-end privacy and security. Existing federated learning-based industrial classification solutions mostly rely on machine learning models for feature extraction and classification decisions. However, there is room for optimization in areas such as lightweight deployment at edge nodes and privacy protection of model interaction information, which need to be adapted to the constraints of industrial environments. Therefore, a privacy protection classification method for industrial big data based on federated learning is proposed to address the aforementioned issues. Summary of the Invention

[0004] The purpose of this invention is to propose a privacy-preserving classification method for industrial big data based on federated learning in order to solve the above-mentioned problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A federated learning-based classification method for privacy protection in industrial big data includes: Multidimensional time series data are collected at various industrial nodes, continuous physical quantities are converted into discrete feature strings, and discrete feature strings are mapped to hash index coordinates; The Bloom filter bit array is initialized locally on the industrial node. The corresponding bit position is set to logic 1 according to the hash index coordinate. The Bloom filter bit array is encrypted bit by bit to generate a ciphertext array. The encrypted array is divided into blocks and an integrity check code is generated and sent to the federation center server. The federation center server receives and reassembles the encrypted array of the active nodes within a preset global aggregation time window. The federated center server aggregates the ciphertext arrays of all active nodes element by element to generate a global ciphertext array. It then extracts the ciphertext elements of the hash index set corresponding to the industrial classification rules from the global ciphertext array, performs homomorphic aggregation, and introduces a blinding factor to generate blinded matching ciphertext. The federal center server calculates the difference between the blinded matching ciphertext and the preset threshold ciphertext, blinds it, and sends it to the trusted decryption node for decryption. The server determines the sign of the difference based on the position of the decryption result in the modular arithmetic space and outputs a Boolean classification result vector to execute industrial control instructions.

[0006] Preferably, the specific process of converting continuous physical quantities into discrete feature strings using an adaptive quantization function is as follows: Obtain the observation values ​​of continuous physical quantities, combine the minimum theoretical physical boundary value that the physical quantity may appear with the quantization step size, and round down to calculate the discrete integer step value. The integer step value is converted into a string form with standard character encoding, and then concatenated with the sensor's unique physical identifier and timestamp to generate a discrete feature string with unique spatiotemporal attributes.

[0007] Preferably, the process of mapping discrete feature strings to hash index coordinates is as follows: The discrete feature string is hashed using a dual hashing technique, employing two independent basic hash functions with different seeds. The results of the two basic hash functions are linearly combined, and a modulo operation is performed on the total length of the Bloom filter bit array to generate multiple hash index coordinates that strictly fall within the valid range.

[0008] Preferably, before initializing the Bloom filter bit array locally at the industrial node, the method further includes: Based on the maximum number of feature strings expected to be inserted into the Bloom filter and the total length of the Bloom filter bit array, the optimal number of hash functions is dynamically calculated and locked with the goal of minimizing the false positive rate. In the initialization state, all bits in the Bloom filter bit array are strictly set to logic zero.

[0009] Preferably, the specific process of encrypting the Bloom filter bit array bit by bit using the additive homomorphic encryption algorithm is as follows: For each bit in the Bloom filter bit array, a random integer coprime to the modulus of the public key is generated using a true random number generator; Determine the logical state of the plaintext bits. When the plaintext bit is logical zero, only calculate the modulus power of the random integer as the ciphertext element. When the plaintext bit is logic 1, the modulus of the random integer is multiplied by the generator to obtain the ciphertext element, thereby avoiding complex modular exponentiation operations on large integers.

[0010] Preferably, the process of dividing the ciphertext array into blocks and generating integrity check codes is as follows: The system obtains real-time network uplink bandwidth and network round-trip latency between industrial nodes and the federation center server, and dynamically calculates the optimal data block size by combining the network congestion control coefficient. The ciphertext array is divided into multiple consecutive ciphertext data blocks according to the optimal data block size; Using a pre-shared key, combined with the strictly increasing sequence number of the ciphertext data block and the precise timestamp, an integrity check code is generated for each ciphertext data block through a hash-based message verification code calculation function.

[0011] Preferably, the process by which the federated center server receives and reassembles the ciphertext array of active nodes within a preset global aggregation time window includes: An independent circular buffer is allocated in memory for each industrial node participating in federated computing; Upon receiving a data packet, the integrity check code is extracted and reverse-checked. If the check is successful, the data is written into the circular buffer in sequence for reassembly. When the preset global aggregation time window is reached, the reception of new ciphertext data blocks is immediately stopped, and it is determined whether the total number of active nodes that have successfully uploaded the complete ciphertext array has reached the preset minimum federation participation threshold. If it has, the ciphertext array of the active nodes is used to enter the global ciphertext rule matching and aggregation step.

[0012] Preferably, the process of introducing a blinding factor to generate blinded matching ciphertext includes: The global ciphertext array is divided into multiple sub-intervals and assigned to different thread pools for parallel modular multiplication and reduction calculations to complete element-by-element aggregation. Extract the ciphertext elements of the hash index set corresponding to the industrial classification rules in the global ciphertext array, and aggregate them into a single matching ciphertext using homomorphic multiplication; A positive integer blinding factor is generated using a true random number generator, and the matched ciphertext is subjected to scalar multiplication homomorphic blinding processing to generate blinded matched ciphertext that masks the absolute frequency value.

[0013] Preferably, the process of calculating the difference between the blinded matching ciphertext and the preset threshold ciphertext and then blinding it is as follows: Obtain the global trigger threshold corresponding to the industrial classification rule, and use the public key to encrypt the global trigger threshold to generate threshold ciphertext; By utilizing the homomorphic subtraction property of the additive homomorphic encryption algorithm, the product of the inverses of the matched ciphertext and the threshold ciphertext in the modular multiplication group is calculated in the ciphertext state to obtain the difference ciphertext. A positive integer blinding factor is generated again to perform scalar multiplication homomorphic blinding on the difference ciphertext, and the blinded difference ciphertext is sent to the trusted decryption node.

[0014] Preferably, the specific process of determining the sign of the difference based on the position of the decryption result in the modular arithmetic space is as follows: The trusted decryption node uses its private key to decrypt the blinded difference ciphertext, obtaining the plaintext difference result that falls within the modulo residual ring interval; Compare the plaintext difference result with half of the modulus; If the plaintext difference result is less than half of the modulus, the original difference is determined to be positive or zero, and the Boolean classification result is output as satisfying the industrial classification rules. If the plaintext difference result is greater than half of the modulus, the original difference is determined to be negative by using the two's complement property of the modulo operation space, and the Boolean classification result is output as not satisfying the industrial classification rules.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention constructs a full-process ciphertext calculation system based on Paillier additive homomorphic encryption, which includes local data feature encryption, global ciphertext aggregation, ciphertext rule matching, and blind differential judgment. The entire process does not require decryption of the original data of any industrial node. At the same time, through a dual cryptographic blinding mechanism and threshold differential judgment logic, the decryption stage only outputs Boolean classification results, without disclosing any global statistical values ​​or local node data information, thus resisting various privacy theft risks such as differential attacks and data restoration attacks.

[0016] 2. This invention achieves a flexible balance between classification accuracy and resource overhead through adaptive quantization discretization. The underlying mathematical optimization of dual hash mapping and homomorphic encryption reduces computational complexity. Combined with a parallel computing architecture, it fully meets the real-time requirements of industrial scenarios. At the same time, through adaptive block transmission and asynchronous time window aggregation mechanism, it can adapt to the complex network environment of industrial IoT, and deal with the global computational blocking problem caused by node transmission delay. The rule-based classification logic can flexibly adapt to the classification needs of different industrial scenarios. Attached Figure Description

[0017] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0019] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0020] Example 1

[0021] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0022] Appendix Figure 1 The flowchart of the industrial big data privacy protection classification method based on federated learning provided in the embodiments of the present invention illustrates the complete steps from collecting multi-dimensional time series data to outputting a Boolean classification result vector to execute industrial control instructions.

[0023] In this embodiment, it includes: Industrial data acquisition and hash mapping involves collecting multidimensional time series data at various industrial nodes, using an adaptive quantization function to convert continuous physical quantities into discrete feature strings, and mapping the discrete feature strings to hash index coordinates through a hash family containing multiple independent hash functions. The core objective of this step is to transform the continuous, multi-dimensional, and heterogeneous low-level sensor physical data generated in industrial settings (such as manufacturing plants, power plants, and other federated participating nodes) into a set of discrete strings that can be received by a Bloom filter through rigorous mathematical transformation and discretization. This string is then mapped to index coordinates in a high-dimensional space using multiple hash functions. No feature extraction networks or machine learning models are involved throughout the process.

[0024] Specifically, this step includes the following extremely detailed operational procedures: First, in each independent industrial federation node (denoted as the node set) ,in To deploy a multi-source sensor network at the underlying device level (total number of industrial nodes participating in federated computing).

[0025] The sensor network operates at a preset fixed sampling frequency. (According to the Nyquist sampling theorem, this frequency must be greater than or equal to twice the highest frequency component of the industrial physical signal to prevent signal aliasing.) Real-time acquisition of operating status data from industrial equipment is required. This data includes, but is not limited to, continuous physical quantities (such as temperature). ,pressure Vibration acceleration ,Voltage ) as well as discrete state quantities (such as valve opening / closing state, relay closing state).

[0026] Assuming within the time window Inside, node The collected raw multidimensional time series data set is denoted as ,in This represents the total number of data sampling points.

[0027] Since a Bloom filter is essentially a probabilistic data structure based on set elements, it cannot directly process continuous floating-point physical data. Therefore, the original continuous data must be rigorously quantized and discretized mathematically.

[0028] For data sets Any continuous physical quantity observation value (For example, the temperature value at a certain moment), the system introduces an adaptive quantization function. This is mapped to discrete integer step values. The quantization function is defined as: ; In this formula, This represents the currently collected continuous physical quantity observation value; This represents the minimum theoretical physical boundary value (e.g., absolute zero or the lowest dormant temperature of the equipment) that this type of physical quantity may appear under normal operation and extreme conditions of industrial equipment. Indicates the quantization step size, i.e., resolution; symbol This indicates the floor function.

[0029] Quantization step size The value of directly determines the classification accuracy and the capacity consumption of the Bloom filter; its calculation formula is as follows: ,in For the maximum theoretical physical boundary value, To quantize bit depth (e.g., take...) The physical interval is then divided into 4096 discrete steps.

[0030] After quantizing all continuous data, the system needs to strictly concatenate the quantized integer values, the sensor's unique physical identifier, and the timestamp into a string to generate a discrete feature string with unique spatiotemporal attributes.

[0031] The splicing rules are defined as follows: in, This represents a string concatenation operation. A unique identifier for the sensor. For timestamps accurate to the millisecond level, This converts quantized integers to their standard ASCII string representation. Therefore, the node... The original dataset Transformed into a set of discrete feature strings .

[0032] The system introduces a containing Hash family with individual hash functions To reduce the CPU clock cycle consumption of hash calculations on industrial edge computing nodes with limited computing resources, this solution abandons computationally intensive cryptographic hashing (such as SHA-3) and instead adopts non-cryptographic high-frequency hashing algorithms based on double hashing technology (such as MurmurHash3 or CityHash).

[0033] Double hashing only requires calculating two basic hash values, which can be generated through a linear combination. Each is an independent pseudo-random hash value.

[0034] The specific mapping formula is as follows: In this formula, Indicates the first A hash function for the feature string The calculated hash index value; The index of the hash function, with a value ranging from integers. ; and These are two independent, base MurmurHash3 functions with different seeds; This represents the total length (i.e., the total number of bits) of the Bloom filter bit array in subsequent steps. Indicates to The modulo operation ensures that the calculated index value strictly falls within the range of 0. Within the closed interval.

[0035] Through the above hash mapping operation, the industrial node Each discrete feature string in They are all deterministically mapped to Located in The range of integer index coordinates. These index coordinates constitute the underlying data feature fingerprint of the industrial node within the current time window, laying a pure mathematical foundation for the subsequent construction of the Bloom filter.

[0036] Local Bloom filter construction and encryption: The Bloom filter bit array is initialized locally on the industrial node. The corresponding bit position is set to logic 1 according to the hash index coordinate. Then, the Bloom filter bit array is encrypted bit by bit using the additive homomorphic encryption algorithm to generate a ciphertext array. The core objective of this step is to construct a Bloom filter bit array representing the characteristics of industrial data locally on the industrial node using the hash index coordinates generated in the previous step, and then introduce an additive homomorphic encryption algorithm to encrypt this bit array bit by bit. This process ensures that after the industrial data is transformed into a feature fingerprint, the fingerprint itself is also protected by a strong cryptographic algorithm, thereby achieving "data usable but not visible" in subsequent federated interactions and preventing the possibility of reverse-engineering the original industrial parameters.

[0037] First, at industrial nodes In local memory, the system initializes a space of length [length missing]. A one-dimensional Bloom filter bit array, denoted as .

[0038] In the initial state, all bits in the array are strictly set to logic 0, that is... .

[0039] Bloom filter length With the number of hash functions The selection of these parameters is not arbitrary, but based on rigorous probability theory derivation to control the false positive rate during classification matching. .

[0040] The mathematical expectation formula for the false positive rate is: in, The maximum number of feature strings to be inserted into the Bloom filter; It is a natural constant.

[0041] In order to in a given and To minimize the false positive rate, the system calculates the optimal number of hash functions by differentiating the above formula and setting the derivative to zero. The theoretical value is: The system determines its memory capacity limit based on the industrial node during the initialization phase. And based on historical data traffic estimates Thus, dynamically calculate and lock value.

[0042] After the in-bit array is initialized, the system iterates through each feature string generated in the previous step. corresponding hash index coordinates .

[0043] For each index coordinate, the system will use a Bloom filter bit array. The corresponding bit in the middle is flipped from logic 0 to logic 1.

[0044] Its mathematical expression is: for all Perform the assignment operation: After analyzing the set Iterate through and insert all elements into the array, which was originally a byte array containing only zeros. It becomes a sparse binary vector containing several 1s and 0s. This binary vector is the "plaintext feature fingerprint" of the industrial node within the current time window.

[0045] To meet the privacy requirements of the federated learning architecture, plaintext bit arrays must be... Encryption is performed. Since subsequent global rule matching requires aggregation and comparison in the ciphertext state, this scheme adopts the Paillier public-key cryptosystem with additive homomorphic properties.

[0046] The Paillier algorithm allows multiplication of two ciphertexts without decryption, and the decrypted result is equivalent to addition of the corresponding plaintexts.

[0047] The key generation and encryption process of the Paillier encryption algorithm is extremely rigorous, as detailed below: Key generation phase: The federal central server (or a trusted third party) randomly selects two large prime numbers of equal and independent length. and (Under industrial-grade safety standards, it is typically required) and The length must be at least 1024 bits to resist existing integer factorization attacks. Calculate the product of these two large prime numbers. and their Carmichael functions ,in This indicates finding the least common multiple.

[0048] Select a generator (i.e., in the model) Selecting from the invertible integer multiplication group To improve computational efficiency, it is usually necessary to directly select... .

[0049] Calculate modular inverse elements , where the function Defined as Finally, a public key is generated. and strictly confidential private keys The federal center will hold the public key. Distributed to all participating industrial nodes private key Securely stored by the federal center.

[0050] Bit-by-bit encryption stage of the bit array: Industrial nodes Received public key Then, the locally generated Bloom filter bit array is processed. Perform bit-by-bit encryption. Because... The length is The node needs to perform This is a separate encryption operation. For bit arrays The first in element (Its value is 0 or 1, where) The system first uses a hardware true random number generator (TRNG) to generate a random integer. ,Require satisfy and and Coprime (i.e., greatest common divisor) ).

[0051] Introducing random numbers The goal is to achieve semantic security, meaning that encrypting the same plaintext (such as multiple 0s or multiple 1s) will produce distinct ciphertexts each time, thus resisting chosen-plaintext attacks (CPA). Subsequently, the system uses the public key to... Encrypt and calculate the ciphertext. The formula is: In this formula, This represents the Paillier encryption function; Represents generator of Power; Represents random numbers of Power; Indicates to Take the mold.

[0052] Due to the length of the Bloom filter in industrial big data Typically extremely large (e.g.) To address the CPU computation bottleneck caused by massive modular exponentiation operations, this solution involves deep optimization of the underlying computing architecture: When plaintext When, the formula simplifies to When plaintext At that time, due to Expanding according to the binomial theorem, Therefore, the formula simplifies to Through this mathematical transformation, the system avoids problems with large integers. Complex modular exponentiation operations reduce computational complexity from Reduced to simply calculating the modulo power of random numbers.

[0053] After the above bit-by-bit encryption operation, the industrial node Successfully converted plaintext bit array Convert to ciphertext array Each element in this ciphertext array is 2048 bits long (assuming...). A large integer (1024 bits).

[0054] At this point, the privacy protection and feature extraction of local data are complete, and the data is ready to enter the federal network for secure interaction.

[0055] The encrypted transmission and synchronization of federated nodes involves dividing the encrypted array into blocks and generating an integrity check code, which is then sent to the federated central server. The federated central server receives and reassembles the encrypted array of the active nodes within a preset global aggregation time window. The core objective of this step is to process the massive ciphertext arrays generated locally at each industrial node. Secure, efficient, and complete data is transmitted to the federated central server. Because the Paillier homomorphic encryption algorithm causes severe data bloat (e.g., 1 bit of plaintext encrypted with a 1024-bit key expands to a 2048-bit ciphertext integer), and Industrial Internet of Things (IIoT) environments typically face severe challenges such as bandwidth constraints, network jitter, and uneven node computing power, this step designs an asynchronous encrypted data synchronization mechanism based on adaptive bandwidth to solve the "fallback" problem caused by network latency in traditional federated architectures.

[0056] Specifically, this step includes the following extremely detailed operational procedures: First, at industrial nodes The system needs to handle lengths of... ciphertext array The system performs low-level serialization and chunking. Since directly encapsulating millions of 2048-bit integers into a single network data packet would cause severe memory overflows and transmission timeouts, the system introduces a dynamic chunking transmission protocol. The system first obtains the current industrial node... Real-time network uplink bandwidth between the federal center server and the server (Unit: Mbps) and round-trip delay (Unit: milliseconds)

[0057] Based on these two network state parameters, the system dynamically calculates the optimal data block size. The calculation formula is: ; In this formula, This is the network congestion control coefficient, and its value range is... The denominator 8 is used to reserve some bandwidth to cope with sudden network jitter; the denominator 8 is used to convert bits into bytes.

[0058] According to the calculation The system will use a large array of ciphertext. Cut into A contiguous block of ciphertext data, denoted as .

[0059] After the data is segmented, in order to ensure the absolute integrity of the encrypted data during transmission in complex industrial networks and to prevent bit flipping caused by man-in-the-middle attacks (MITM) or channel noise, the system generates a strict cryptographic checksum for each encrypted data block.

[0060] The system uses the SHA-256 hash algorithm, combined with industrial nodes. pre-shared key Calculate the hash-based message verification code (HMAC). For the first... Data blocks (in ), its integrity check code The calculation formula is: ; In this formula, This represents the HMAC calculation function based on SHA-256; This indicates the concatenation operation of the underlying byte stream; This is a strictly increasing sequence number for the current data block, used to prevent replay attacks. This is the precise timestamp when the data block was generated.

[0061] Subsequently, industrial nodes A two-way authenticated encrypted tunnel is established with the federated center server via Transport Layer Security (TLS 1.3). The node will encapsulate the data packets (including block sequence numbers)... ciphertext data blocks and verification code It is sent to the central server through an asynchronous non-blocking I / O multiplexing mechanism (such as epoll or kqueue).

[0062] On the federated center server side, in order to handle the high concurrency connections from a massive number of industrial nodes, the server stores information in memory for each node participating in federated computation. An independent circular buffer zone was created.

[0063] After the server receives the data packet, it first extracts... Use the corresponding Perform a reverse verification. If the verification fails, discard the data packet and request the node to retransmit; if the verification succeeds, then... According to the serial number Write sequentially to the circular buffer.

[0064] When node All After all data blocks were successfully received and reassembled, the server reconstructed the complete ciphertext array in memory. .

[0065] To address the common "fallback" problem in industrial networks, this solution abandons the traditional synchronous federation waiting mechanism (which requires all nodes to complete uploading before proceeding to the next step) and instead adopts an asynchronous truncation strategy based on time windows. The federation center sets a strict global aggregation time window. When the time comes... At this point, the server immediately stops receiving new data blocks. Let the set of nodes that successfully uploaded the complete ciphertext array at this time be denoted as . (in And its number of elements is denoted as ,satisfy ).if only Reaching the preset minimum federal participation threshold (For example The server immediately utilizes the collection The encrypted data in the middle enters the next global aggregation stage; it failed to... Data transmitted within the node will be discarded directly.

[0066] Global ciphertext rule matching and aggregation: The federated center server uses homomorphic multiplication to aggregate the ciphertext arrays of all active nodes element by element to generate a global ciphertext array. It then extracts the ciphertext elements of the hash index set corresponding to the industrial classification rule in the global ciphertext array and performs homomorphic aggregation. Finally, it introduces a blinding factor to generate blinded matching ciphertext.

[0067] The core objective of this step is to ensure that the federal center server does not possess Paillier's private key. Under the premise of absolute zero knowledge that cannot decrypt any node's data, the mathematical properties of homomorphic encryption are used to globally aggregate the ciphertext Bloom filters uploaded by all active nodes and perform pure mathematical matching in the ciphertext state with pre-set industrial classification rules.

[0068] This step transforms the complex industrial classification logic into underlying large integer modular exponentiation and modular multiplication operations.

[0069] First, the federal center server needs to process the received data. Perform global homomorphic aggregation on a ciphertext array. Assume a set of active nodes. The uploaded ciphertext array set is The server initializes a memory array of the same length. global ciphertext array According to the additive homomorphism property of the Paillier cryptosystem, the product of two ciphertexts, after decryption, is equal to the sum of their corresponding plaintexts.

[0070] Therefore, in order to obtain the global frequency statistics of all nodes at each bit of the Bloom filter, the server needs to perform element-wise homomorphic multiplication on all ciphertext arrays. For the th bit in the global ciphertext array... element (in The mathematical formula for homomorphic aggregation is: ; In this formula, This indicates a series of multiplication operations; Indicates the first The ciphertext array uploaded by the active node is the first one. One encrypted element; The square of the modulus in the Paillier public key; Indicates in the model Multiplication is performed in the remaining class ring.

[0071] To prove the correctness of this aggregation operation from a rigorous mathematical perspective, we will... Substituting the encryption formula into the above multiplication formula, we can derive the following: According to the associative and commutative laws of exponentiation, the above equation can be simplified to: This will increase the global plaintext frequency. Global random numbers .

[0072] Since the product of multiple coprime random numbers is still... Coprime, therefore It is still a valid Paillier random number.

[0073] The final derivation result is: ; The above mathematical derivation proves that the multiplication operation performed by the server in encrypted state is perfectly equivalent to the bitwise accumulation of the plaintext Bloom filter bit groups at the bottom layer of all industrial nodes.

[0074] After aggregation Essentially, it's a count Bloom filter in an encrypted state, hiding the plaintext. Indicates the first The total number of times the hash index is triggered across all nodes.

[0075] because The scale of these data is extremely large (in the millions), and modular multiplication of large integers is extremely CPU clock cycles-intensive. To meet the real-time requirements of industrial big data, the server employs a Single Instruction Multiple Data (SIMD) instruction set (such as AVX-512) and a multi-threaded concurrent computing model in its underlying computing architecture. The server will handle data of length... The array is divided into Sub-intervals ( (The number of logical cores of the server CPU) is allocated to different thread pools for parallel modular multiplication and reduction calculations, thereby reducing the aggregation time complexity from... Reduce to .

[0076] After completing global ciphertext aggregation, the system enters the ciphertext rule matching phase. The federated center pre-maintains an industrial classification rule base. ,in This represents the total number of classification rules. Each rule... (For example, "spindle bearing temperature is too high and vibration frequency is abnormal") has been converted into a standard discrete feature string during system initialization and uses the same hash family as in step 1. Mapped to A set of hash index coordinates, denoted as index set. .

[0077] To determine whether the rules are satisfied in the global data under encrypted state. The server needs to extract the global ciphertext array. The middle corresponds to the index set of Each ciphertext element is then combined using homomorphic addition (i.e., ciphertext multiplication). Each ciphertext element is aggregated into a single matching ciphertext. .

[0078] Match ciphertext The calculation formula is: ; The physical significance of this formula lies in the fact that it calculates the rules in the encrypted state. Corresponding The sum of global frequencies at each hash position. If this The extremely high sum of the frequencies at each location indicates that the industrial fault rule is frequently triggered across multiple federated nodes.

[0079] However, if directly The message is sent to the decryptor who holds the private key. The decryptor will then directly obtain the exact sum of frequencies, which in extreme cases could lead to differential attacks targeting specific nodes.

[0080] To close this privacy vulnerability and meet the highest level of security requirements, the server outputs... Previously, a cryptographic blinding operation had to be introduced. The server used a true random number generator to generate a sufficiently large positive integer blinding factor. (Require The ciphertext must be at least 64 bits long, and Paillier's scalar multiplication homomorphism property (i.e., the power of the plaintext in the ciphertext equals the product of the plaintext) is used to... Perform blinding processing.

[0081] The final matching ciphertext after blinding The calculation formula is: ; Based on the homomorphic property, this operation is equivalent to multiplying the sum of frequencies by an unknown random number at the plaintext level. .

[0082] That is, the value obtained after decryption will be Through this ingenious underlying mathematical transformation, the server preserves the relative magnitude of the sum of frequencies (for subsequent threshold determination) while masking the absolute frequency values, thus achieving perfect privacy protection.

[0083] The server iterates through all the rules in the rule base. Each rule is used to calculate the corresponding blinded matching ciphertext set in turn. At this point, the global ciphertext rule matching and aggregation steps are complete, and the system will... Submit to the next step for final privacy decryption and classification determination.

[0084] Privacy decryption and industrial classification determination: The federal center server calculates the difference between the blinded matching ciphertext and the preset threshold ciphertext and performs blinding, then sends it to the trusted decryption node for decryption. The sign of the difference is determined based on the position of the decryption result in the modular operation space, and a Boolean classification result vector is output to execute industrial control instructions. The core purpose of this step is to decrypt the blinded ciphertext by a trusted third party (or the secure and isolated execution environment of the federation center) holding the Paillier private key, without disclosing any single industrial node's underlying raw data or global accurate frequency statistics. Then, through the positive and negative judgment logic in the modular operation space, the final industrial classification result is obtained, thereby triggering the corresponding industrial control instructions.

[0085] In the previous stage, although the federal center server calculated the rules... Corresponding blind matching ciphertext However, if the ciphertext is decrypted directly, the decryptor will still obtain a blinding factor. Amplified frequency values. In order to achieve "outputting only Boolean classification results (yes / no rule satisfied) without outputting any specific numerical values" at a purely mathematical level, this scheme introduces a ciphertext threshold differential judgment mechanism before decryption.

[0086] Specifically, for each rule in the industrial classification rule base The industrial expert system has a pre-set global trigger threshold. (For example, at least 30% of the nodes must trigger this abnormal feature before it is classified as a global systemic failure.)

[0087] Since each rule is mapped to Therefore, the theoretical minimum total trigger frequency for this rule in the global Bloom filter should be [number of hash indices]. .

[0088] Federal center servers use public keys Encrypt the plaintext threshold to generate threshold ciphertext. : in, This is a newly generated random number.

[0089] Subsequently, the server utilizes the homomorphic subtraction property of the Paillier cryptosystem (i.e., ciphertext multiplication combined with modular inverse operation) to calculate the globally matching ciphertext in the ciphertext state. With threshold ciphertext The difference ciphertext is denoted as .

[0090] The calculation formula is: In this formula, Indicates the threshold ciphertext in the model The inverse element in a multiplicative group. Based on homomorphic properties, the plaintext corresponding to this difference ciphertext after decryption is: the sum of the actual global frequencies minus the theoretical trigger threshold (i.e., ... ).

[0091] To prevent the decryptor from knowing the exact difference, the server again uses a true random number generator to generate a positive integer blinding factor with a length of at least 64 bits. The difference ciphertext is homomorphically blinded using scalar multiplication to generate the final blinded difference ciphertext. : ; The server will The set of blinded difference ciphertexts corresponding to each rule Send to the holder of the private key A trusted decryption node.

[0092] After receiving the blinded differential ciphertext, the trusted decryption node initiates the Paillier core decryption algorithm. For any given... The decryption formula is: ; In this formula, ; The value of the Carmichael function; For modular inversion elements; This is the plaintext result obtained after decryption.

[0093] Classification and determination logic (based on the two's complement property of the modulo operation space): due to the decryption result In the model It is calculated from the remaining class ring, and it must be a ring that falls within the ring of classes. Positive integers within the interval. However, the actual frequency difference may be negative (i.e., the global frequency has not reached the threshold). In the modulus... In arithmetic, negative numbers It is represented in its two's complement form. .

[0094] because It is a 2048-bit large integer (its value is extremely large, far exceeding the frequency of any possible industrial data), and the system can determine this by... The sign of the difference is determined by its position in the modulus space. if If the original difference is positive or zero, it means the actual global frequency is greater than or equal to the trigger threshold, according to this industrial classification rule. The condition is satisfied (determined as logic 1); if This indicates that the original difference was negative (actually...). (A smaller absolute value was subtracted). This means the actual global frequency is less than the trigger threshold, according to the industrial classification rule. Not satisfied (determined as logic 0).

[0095] Based on the above determination, the decryption node transforms the complex ciphertext set into a string of length [length missing]. Boolean classification result vector ,in Finally, the decryption node inputs the Boolean vector into the pre-defined industrial state decision tree. For example, if the rule... (Excessive temperature) and If both (vibration exceeding the standard) and 1 are true, the system will output the classification label "severe spindle wear failure"; If all rules are 0, the "System operating normally" category label is output. The decryption node broadcasts the final category label to each industrial node through an encrypted channel. The programmable logic controller (PLC) at the bottom of the industrial node automatically executes the corresponding physical control actions (such as emergency shutdown, starting the backup cooling pump, etc.) according to the category label.

[0096] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0097] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0098] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0099] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0100] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A privacy-preserving classification method for industrial big data based on federated learning, characterized in that, include: Multidimensional time series data are collected at various industrial nodes, continuous physical quantities are converted into discrete feature strings, and discrete feature strings are mapped to hash index coordinates; The Bloom filter bit array is initialized locally on the industrial node. The corresponding bit position is set to logic 1 according to the hash index coordinate. The Bloom filter bit array is encrypted bit by bit to generate a ciphertext array. The encrypted array is divided into blocks and an integrity check code is generated and sent to the federation center server. The federation center server receives and reassembles the encrypted array of the active nodes within a preset global aggregation time window. The federated center server aggregates the ciphertext arrays of all active nodes element by element to generate a global ciphertext array. It then extracts the ciphertext elements of the hash index set corresponding to the industrial classification rules from the global ciphertext array, performs homomorphic aggregation, and introduces a blinding factor to generate blinded matching ciphertext. The federal center server calculates the difference between the blinded matching ciphertext and the preset threshold ciphertext, blinds it, and sends it to the trusted decryption node for decryption. The server determines the sign of the difference based on the position of the decryption result in the modular arithmetic space and outputs a Boolean classification result vector to execute industrial control instructions.

2. The industrial big data privacy protection classification method based on federated learning according to claim 1, characterized in that, The specific process of converting a continuous physical quantity into a discrete feature string using an adaptive quantization function is as follows: Obtain the observation values ​​of continuous physical quantities, combine the minimum theoretical physical boundary value that the physical quantity may appear with the quantization step size, and round down to calculate the discrete integer step value. The integer step value is converted into a string form with standard character encoding, and then concatenated with the sensor's unique physical identifier and timestamp to generate a discrete feature string with unique spatiotemporal attributes.

3. The industrial big data privacy protection classification method based on federated learning according to claim 2, characterized in that, The process of mapping discrete feature strings to hash index coordinates is as follows: The discrete feature string is hashed using a dual hashing technique, employing two independent basic hash functions with different seeds. The results of the two basic hash functions are linearly combined, and a modulo operation is performed on the total length of the Bloom filter bit array to generate multiple hash index coordinates that strictly fall within the valid range.

4. The industrial big data privacy protection classification method based on federated learning according to claim 1, characterized in that, Before the industrial node locally initializes the Bloom filter bit array, the following is also included: Based on the maximum number of feature strings expected to be inserted into the Bloom filter and the total length of the Bloom filter bit array, the optimal number of hash functions is dynamically calculated and locked with the goal of minimizing the false positive rate. In the initialization state, all bits in the Bloom filter bit array are strictly set to logic zero.

5. The industrial big data privacy protection classification method based on federated learning according to claim 4, characterized in that, The specific process of encrypting the bit array of the Bloom filter bitwise using the additive homomorphic encryption algorithm is as follows: For each bit in the Bloom filter bit array, a random integer coprime to the modulus of the public key is generated using a true random number generator; Determine the logical state of the plaintext bits. When the plaintext bit is logical zero, only calculate the modulus power of the random integer as the ciphertext element. When the plaintext bit is logic 1, the modulus of the random integer is multiplied by the generator to obtain the ciphertext element, thereby avoiding complex modular exponentiation operations on large integers.

6. The industrial big data privacy protection classification method based on federated learning according to claim 1, characterized in that, The process of dividing the ciphertext array into blocks and generating integrity check codes is as follows: The system obtains real-time network uplink bandwidth and network round-trip latency between industrial nodes and the federation center server, and dynamically calculates the optimal data block size by combining the network congestion control coefficient. The ciphertext array is divided into multiple consecutive ciphertext data blocks according to the optimal data block size; Using a pre-shared key, combined with the strictly increasing sequence number of the ciphertext data block and the precise timestamp, an integrity check code is generated for each ciphertext data block through a hash-based message verification code calculation function.

7. The industrial big data privacy protection classification method based on federated learning according to claim 6, characterized in that, The process by which the federated center server receives and reassembles the ciphertext array of active nodes within a preset global aggregation time window includes: An independent circular buffer is allocated in memory for each industrial node participating in federated computing; Upon receiving a data packet, the integrity check code is extracted and reverse-checked. If the check is successful, the data is written into the circular buffer in sequence for reassembly. When the preset global aggregation time window is reached, the reception of new ciphertext data blocks is immediately stopped, and it is determined whether the total number of active nodes that have successfully uploaded the complete ciphertext array has reached the preset minimum federation participation threshold. If it has, the ciphertext array of the active nodes is used to enter the global ciphertext rule matching and aggregation step.

8. The industrial big data privacy protection classification method based on federated learning according to claim 1, characterized in that, The process of introducing a blinding factor to generate blinded matching ciphertext includes: The global ciphertext array is divided into multiple sub-intervals and assigned to different thread pools for parallel modular multiplication and reduction calculations to complete element-by-element aggregation. Extract the ciphertext elements of the hash index set corresponding to the industrial classification rules in the global ciphertext array, and aggregate them into a single matching ciphertext using homomorphic multiplication; A positive integer blinding factor is generated using a true random number generator, and the matched ciphertext is subjected to scalar multiplication homomorphic blinding processing to generate blinded matched ciphertext that masks the absolute frequency value.

9. The industrial big data privacy protection classification method based on federated learning according to claim 1, characterized in that, The process of calculating the difference between the blinded matching ciphertext and the preset threshold ciphertext and then blinding it is as follows: Obtain the global trigger threshold corresponding to the industrial classification rule, and use the public key to encrypt the global trigger threshold to generate threshold ciphertext; By utilizing the homomorphic subtraction property of the additive homomorphic encryption algorithm, the product of the inverses of the matched ciphertext and the threshold ciphertext in the modular multiplication group is calculated in the ciphertext state to obtain the difference ciphertext. A positive integer blinding factor is generated again to perform scalar multiplication homomorphic blinding on the difference ciphertext, and the blinded difference ciphertext is sent to the trusted decryption node.

10. The industrial big data privacy protection classification method based on federated learning according to claim 9, characterized in that, The specific process for determining the sign of the difference based on the position of the decryption result in the modular arithmetic space is as follows: The trusted decryption node uses its private key to decrypt the blinded difference ciphertext, obtaining the plaintext difference result that falls within the modulo residual ring interval; Compare the plaintext difference result with half of the modulus; If the plaintext difference result is less than half of the modulus, the original difference is determined to be positive or zero, and the Boolean classification result is output as satisfying the industrial classification rules. If the plaintext difference result is greater than half of the modulus, the original difference is determined to be negative by using the two's complement property of the modulo operation space, and the Boolean classification result is output as not satisfying the industrial classification rules.

Images