Privacy protection distributed federated learning collaboration strategy for networked automobiles

Abstract

The invention relates to the technical field of privacy protection of networked automobiles, and discloses a privacy protection distributed federated learning collaborative strategy for networked automobiles, which comprises a privacy protection system of networked automobiles, and the privacy protection system of networked automobiles comprises an application scene layer, a core technology layer, an edge computing layer and a data and equipment layer. According to the privacy protection distributed federal learning cooperation strategy for the networked automobile, real-time and multi-dimensional perception of a traffic environment is realized through integration of V2X communication, a road condition space-time encoder, a stream data processing engine and other technologies, accurate data support is provided for intelligent driving decision making, and the privacy protection distributed federal learning cooperation strategy for the networked automobile is provided for the intelligent driving decision making. Through a state division domain strategy module and a time-space domain adaptive division technology, resource allocation and task scheduling can be dynamically adjusted according to real-time road conditions and network conditions, the response speed and efficiency of the system are improved, and through technologies of a privacy calculation pool module, dynamic privacy budget allocation, a PKI certificate service center and the like, real-time dynamic privacy budget allocation can be realized. And strict privacy protection is realized during data sharing and federal learning.

Description

Technical Field

[0001] The present invention relates to the field of privacy protection technology for connected vehicles, and specifically to a privacy-preserving distributed federated learning collaborative strategy for connected vehicles. Background Art

[0002] A connected car is an intelligent car that uses Internet technology to achieve real-time information exchange between vehicles and between vehicles and the external environment. Connected cars use advanced perception, communication and control technologies to achieve information sharing and mutual coordination between vehicles, enabling cars to drive more intelligently and safely, providing drivers with a more convenient and comfortable travel experience. Features of connected cars include autonomous driving, vehicle-to-vehicle communication, vehicle-to-infrastructure communication, in-vehicle entertainment and information services, etc. Through autonomous driving technology, connected cars can achieve functions such as automatic cruising, automatic parking, and automatic obstacle avoidance, greatly reducing driving risks and improving driving efficiency. At the same time, connected cars can also achieve fleet coordinated driving and avoid traffic congestion through vehicle-to-vehicle communication, obtain real-time road conditions information through vehicle-to-infrastructure communication, and plan the best driving route in advance. In addition, connected cars can also provide a variety of in-vehicle entertainment and information services, such as smart navigation, online music, remote control and other functions, so that drivers can enjoy more convenience during driving.

[0003] The privacy protection of connected cars is the protection of personal privacy data, which usually includes the vehicle's location, driving trajectory, vehicle status and driving behavior. Since the Internet of Vehicles system involves a large amount of personal sensitive data, privacy protection has become an important issue in this field. With the development of technology, the privacy protection of Internet of Vehicles data will pay more attention to personalized needs and provide users with more flexible privacy protection solutions.

[0004] Current privacy protection strategies for connected vehicles still have shortcomings in use. For example, 1. The dynamic changes in network topology caused by high-speed vehicle movement are not compatible with traditional static privacy strategies; 2. There is a lack of a unified trust anchor and responsibility tracing mechanism among multiple entities such as vehicles, roads, clouds, and networks; 3. The privacy exposure risks brought about by the aggregation of multi-source data. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In response to the shortcomings of the existing technology, the present invention provides a privacy-preserving distributed federated learning collaborative strategy for connected vehicles, which has the advantages of spatiotemporal collaboration, layered privacy protection and dynamic federated learning. It solves the shortcomings of the current privacy protection strategies for connected vehicles in use, such as the dynamic changes in network topology caused by high-speed movement of vehicles, which are insufficiently adapted to traditional static privacy strategies; the lack of a unified trust anchor point and responsibility tracing mechanism among multiple subjects such as vehicles, roads, clouds and networks; and the privacy exposure risks caused by the aggregation of multi-source data.

[0007] (2) Technical solution

[0008] To achieve the above-mentioned spatiotemporal collaboration, layered privacy protection, and dynamic federated learning, the present invention provides the following technical solutions: a privacy-preserving distributed federated learning collaborative strategy for connected vehicles, including a connected vehicle privacy protection system, wherein the connected vehicle privacy protection system includes an application scenario layer, a core technology layer, an edge computing layer, and a data and device layer;

[0009] The application scenario layer includes: intelligent driving decision-making, communication traffic prediction, V2X communication, multi-dimensional situational awareness and data lineage tracking dashboard;

[0010] The core technology layer includes: dynamic domain strategy module, heterogeneous network adaptation module, privacy computing pool module, federated learning module, incentive and audit module and multi-active disaster recovery scheduler;

[0011] The edge computing layer includes: a road condition spatiotemporal encoder, a stream data processing engine, an edge model lightweight container, and a federated data directory service;

[0012] The data and device layer includes: vehicle-mounted terminal group, OBU equipment, RSU edge node, cloud server, third-party data platform, supervision node, PKI certificate service center and data sensitivity label manager.

[0013] Furthermore, the dynamic domain division strategy module includes adaptive division of spatiotemporal domains, data sensitivity classification and cross-domain coordination mechanism;

[0014] The spatiotemporal domain adaptive partitioning is provided with a spatiotemporal indexing engine for dynamically constructing a three-level domain model of "vehicle terminal-roadside unit-cloud" based on the vehicle's driving trajectory and road topology. For example, a long chain domain is formed in a highway scenario, and a star domain is formed at a city intersection.

[0015] The data sensitivity classification includes sensitive units, semi-sensitive units and public units. The sensitive units contain driving behavior data (braking frequency, steering wheel angle), the semi-sensitive units contain environmental perception data (road condition images, radar point clouds), and the public units contain traffic regulation data (traffic light status and speed limit signs).

[0016] The cross-domain collaboration mechanism can realize the trusted transmission of cross-domain model parameters through blockchain smart contracts and adopt homomorphic encryption technology to ensure the security of inter-domain communication.

[0017] Furthermore, the heterogeneous network adaptation module includes dynamic switching of communication protocols and network slice resource scheduling;

[0018] The dynamic switching of communication protocols includes 5G high-bandwidth environments, V2X low-latency environments, and weak edge network connections;

[0019] The network slice resource scheduling allocates dedicated network slices for federated learning tasks and achieves precise control of bandwidth and latency through software-defined networking (SDN).

[0020] Furthermore, the privacy computing pool module includes a hybrid privacy technology stack and dynamic privacy budget allocation;

[0021] The hybrid privacy technology stack includes a data preprocessing unit, a calculation execution unit, and a result verification unit. The data preprocessing unit is used to add noise to the original data using differential privacy (DP). The calculation execution unit is used to implement joint modeling using secure multi-party computation (MPC). The result verification unit is used to verify the legitimacy of the model using zero-knowledge proof (ZKP).

[0022] The dynamic privacy budget allocation is used to dynamically adjust the privacy budget according to data sensitivity and task importance. For example, 80% of the budget is allocated to steering wheel angle data and 20% is allocated to weather data. A performance monitor and a compliance evaluator are also set up. The performance monitor is used to calculate the accuracy loss in real time, and the compliance evaluator is used to connect to the supervision node.

[0023] Furthermore, the federated learning module includes cross-modal federated learning, incremental model updating, and adaptive aggregation strategy;

[0024] The cross-modal federated learning is provided with a modal feature converter and a heterogeneous gradient compensator. The modal feature converter is used to unify the feature space of multi-source heterogeneous data of the Internet of Vehicles. In the Internet of Vehicles scenario, vehicles will collect images (cameras), point clouds (lidar), millimeter-wave radar data, vehicle status data of the CAN bus, etc. These data modalities are very different. Through the converter, the visual features of the image, the three-dimensional spatial features of the point cloud, etc. can be mapped to a unified feature dimension, so that different modal data can participate in joint training. The heterogeneous gradient compensator is used to compensate for the training deviation caused by the heterogeneity of Internet of Vehicles devices. When the on-board terminal (weak computing power and little data) and the roadside unit (strong computing power and rich data) are trained, the gradient update amplitude and direction are very different. The compensator enhances / corrects the weak device gradient by analyzing the metadata such as the device computing power and data quality to ensure the convergence of the federated training.

[0025] The incremental model update is provided with a model difference analyzer and progressive weight clipping. The model difference analyzer is used to monitor the impact of changes in the Internet of Vehicles environment on the model. When the connected car is driving, the traffic scene (such as morning and evening rush hours, rainy and snowy weather) and road topology (urban road / highway switching) change dynamically, and the data distribution continues to drift. The analyzer compares the prediction difference between the global model and the local model on the new data, identifies the scene change trigger point, and decides whether to start the model update. The progressive weight clipping is used to efficiently update the model while controlling the communication overhead. The communication bandwidth between Internet of Vehicles devices is limited (such as delay fluctuations in V2X communication). Redundant weights (such as model parameters that are insensitive to traffic scene changes) are clipped, and only key gradients / parameter updates are transmitted. While ensuring model accuracy, the amount of transmitted data is reduced to adapt to real-time requirements;

[0026] The adaptive aggregation strategy is used to dynamically select aggregation rules based on the real-time status of the Internet of Vehicles (device online rate, network latency, and data privacy level).

[0027] Furthermore, the incentive and audit module includes a contribution quantification model, a traceable security audit, an abnormal behavior detector, and a risk heat map;

[0028] The contribution quantification model is used to integrate the three dimensions of data volume, data quality, and computing resource consumption, and motivate participants through a reputation points system;

[0029] The traceable security audit uses the consortium chain to record the entire life cycle of the model training, supporting regulatory nodes to verify the compliance of privacy protection measures;

[0030] The abnormal behavior detector is used to monitor anomalies in real time, "draw red lines" for the behavior of participating nodes, and use machine learning models to learn normal participation patterns (data distribution, gradient change patterns, and communication frequencies). Once a node uploads malicious data (fake road conditions, distorted driving behavior), conducts attack training (poisoning the global model to invalidate it), or illegally steals privacy (reversely derives sensitive information from other nodes), for example, "the gradient of a certain vehicle-mounted terminal suddenly becomes abnormally large, which is inconsistent with the data patterns of the vehicle model," an alarm will be triggered immediately, and the risk will be eliminated at the source: before the malicious behavior affects the global model, the "problem node" will be intercepted in advance, isolated, and verified before deciding whether to re-access it, thereby ensuring the model and data security of federated learning.

[0031] The risk heat map is used to visualize risk distribution, marking the node locations (urban sections, regions), risk types (data leakage, model poisoning, device intrusion), and risk levels (high / medium / low) in the Internet of Vehicles on the GIS map. For example, "XX elevated bridge section, 3 vehicle-mounted terminal data are abnormal, and the risk of model poisoning is high", intuitively presenting the risk "hard-hit areas" and accurately allocating resources. By looking at the heat map, car companies and traffic management departments can strengthen supervision in a targeted manner (add security verification and encryption levels to nodes in high-risk areas), optimize privacy policies (adjust the federated learning aggregation frequency / algorithm in risk-concentrated areas), and use limited resources to build a strong regional risk defense line.

[0032] Furthermore, the intelligent driving decision-making is used to make autonomous driving "smart and safe", privacy protection is integrated into the driving decision-making process, the traffic flow prediction uses privacy protection technology, aggregates multi-vehicle data for traffic analysis, and serves smart cities, the V2X communication is used to ensure privacy security during vehicle-to-vehicle and vehicle-to-road (V2V, V2R) communications, the multi-dimensional situational awareness is used to enable the Internet of Vehicles to "see global risks", and privacy protection and security monitoring are both taken into account, and the data lineage tracking dashboard is used to "record the family tree" of Internet of Vehicles data and trace the privacy processing status throughout the entire process.

[0033] Furthermore, the road condition spatiotemporal encoder is used to "standardize and privacy encode" the data, turning the "messy road condition data" in the Internet of Vehicles into a format that can be used to train models while protecting privacy. The stream data processing engine is used to "screen data and resist interference" in real time to cope with the "tide of real-time data" in the Internet of Vehicles, quickly process, denoise, and select useful information. The edge model lightweight container is used to allow the model to "run at the edge without getting stuck", and to split the "big model" of federated learning into a lightweight version that can run on edge devices (on-board terminals, RSUs) to achieve real-time reasoning + local training. The federated data directory service is used to "map and manage permissions" for the data, and to perform "directory indexing and permission management" for the data participating in federated learning on the edge side, so that the data can be "found, used correctly, and not exceed authority."

[0034] Furthermore, the on-board terminal group is the "data collector + edge node" of the connected car, the OBU device is the on-board unit, the "connector" for communication between the car and the outside world, and the main feature is V2X communication (car to car, car to road). The RSU edge node is the roadside unit, the "edge brain" of the Internet of Vehicles, performing real-time data processing + local collaboration on the roadside. The cloud server is the "central brain" of the connected car, performing global model training + big data analysis. The third-party data platform is the "external aid" for supplementary data, providing related data outside the Internet of Vehicles (such as weather and map updates). The supervision node is the "inspector" of privacy protection, supervising data flow + compliance throughout the process. The PKI certificate service center is the "ID card center" of network communication, issuing digital certificates for identity authentication. The data sensitivity label manager is used to "label and grade" data to guide privacy protection strategies.

[0035] (3) Beneficial effects

[0036] Compared with the existing technology, this invention provides a privacy-preserving distributed federated learning collaborative strategy for connected cars, which has the following beneficial effects:

[0037] 1. This privacy-preserving distributed federated learning collaborative strategy for connected vehicles achieves real-time, multi-dimensional perception of the traffic environment by integrating technologies such as V2X communication, road condition spatiotemporal encoders, and streaming data processing engines, providing accurate data support for intelligent driving decisions. Through the state domain strategy module and spatiotemporal domain adaptive partitioning technology, it can dynamically adjust resource allocation and task scheduling according to real-time road and network conditions, improving system response speed and efficiency. Through the privacy computing pool module, dynamic privacy budget allocation, PKI certificate service center and other technologies, the system implements strict privacy protection during data sharing and federated learning to ensure compliance with regulatory requirements.

[0038] 2. This privacy-preserving distributed federated learning collaborative strategy for connected vehicles supports the efficient fusion and collaborative training of multi-source data (such as visual models and radar point cloud models) through technologies such as cross-modal federated learning and heterogeneous gradient compensators, thereby improving model generalization capabilities. It optimizes the computing resource utilization of edge nodes and reduces latency through lightweight edge model containers and heterogeneous network adaptation modules, making it suitable for the deployment requirements of on-board terminals and OBU / RSU equipment. It ensures the traceability of data flow and model training processes through functions such as data lineage tracking and review, contribution quantification models, and abnormal behavior detectors, facilitating supervision and auditing.

[0039] 3. This privacy-preserving distributed federated learning collaborative strategy for connected vehicles achieves dynamic optimization of computing and communication resources through technologies such as network slicing resource scheduling and multi-active disaster recovery schedulers, ensuring system stability under high load or network fluctuations. Through cross-modal alignment layers and multi-sensor fusion reasoning, the system can comprehensively utilize heterogeneous data such as vision and radar point clouds to improve the accuracy and robustness of environmental perception, thereby enabling the system to build an intelligent driving solution covering the entire chain of "data collection-edge processing-collaborative learning-privacy protection-application decision-making". It is real-time, secure, adaptable and scalable, and is suitable for complex Internet of Vehicles environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the overall strategy of the present invention;

[0041] Figure 2 This is a schematic diagram of the core technology layer of the present invention;

[0042] Figure 3 This is a schematic diagram of the spatiotemporal coordination of the present invention;

[0043] Figure 4 Schematic diagram of federated learning in the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] See also Figure 1-4 , a privacy-preserving distributed federated learning collaborative strategy for connected cars, including a privacy protection system for connected cars. The privacy protection system for connected cars includes an application scenario layer, a core technology layer, an edge computing layer, and a data and device layer;

[0046] The application scenario layer includes: intelligent driving decision-making, communication traffic prediction, V2X communication, multi-dimensional situational awareness, and data lineage tracking dashboard;

[0047] The core technology layer includes: dynamic domain strategy module, heterogeneous network adaptation module, privacy computing pool module, federated learning module, incentive and audit module and multi-active disaster recovery scheduler;

[0048] The edge computing layer includes: a road condition spatiotemporal encoder, a stream data processing engine, an edge model lightweight container, and a federated data directory service;

[0049] The data and device layer includes: vehicle-mounted terminal groups, OBU devices, RSU edge nodes, cloud servers, third-party data platforms, supervision nodes, PKI certificate service centers, and data sensitivity label managers.

[0050] In the case implementation, the dynamic domain division strategy module includes adaptive division of spatiotemporal domains, data sensitivity classification, and cross-domain coordination mechanism;

[0051] The spatiotemporal domain adaptive partitioning uses a spatiotemporal indexing engine to dynamically construct a three-level domain model of "vehicle terminal-roadside unit-cloud" based on vehicle trajectories and road topology. For example, a long chain domain is formed in highway scenarios, and a star-shaped domain is formed at urban intersections.

[0052] Data sensitivity levels include sensitive units, semi-sensitive units, and public units. Sensitive units include driving behavior data (braking frequency, steering wheel angle), semi-sensitive units include environmental perception data (road condition images, radar point clouds), and public units include traffic rules data (signal light status and speed limit signs).

[0053] The cross-domain collaboration mechanism can realize the trusted transmission of cross-domain model parameters through blockchain smart contracts and use homomorphic encryption technology to ensure the security of inter-domain communications.

[0054] In the case implementation, the heterogeneous network adaptation module includes dynamic switching of communication protocols and network slice resource scheduling;

[0055] Dynamic switching of communication protocols includes 5G high-bandwidth environments, V2X low-latency environments, and weak edge network connections, for example:

[0056]

[0057] Network slice resource scheduling allocates dedicated network slices for federated learning tasks and achieves precise control of bandwidth and latency through software-defined networking (SDN).

[0058] In the case implementation, the privacy computing pool module includes a hybrid privacy technology stack and dynamic privacy budget allocation;

[0059] The hybrid privacy technology stack includes a data preprocessing unit, a computation execution unit, and a result verification unit. The data preprocessing unit is used for differential privacy (DP) to add noise to the original data. The computation execution unit is used for secure multi-party computation (MPC) to implement joint modeling. The result verification unit is used for zero-knowledge proof (ZKP) to verify the legitimacy of the model. For example:

[0060]

[0061] Dynamic privacy budget allocation is used to dynamically adjust the privacy budget based on data sensitivity and task importance. For example, 80% of the budget is allocated to steering wheel angle data and 20% to weather data. A performance monitor and a compliance evaluator are also set up. The performance monitor is used to calculate accuracy loss in real time, and the compliance evaluator is used to connect to the supervision node.

[0062] The typical allocation strategies for dynamic privacy budget allocation are as follows:

[0063] Highly sensitive data (such as steering wheel angle): allocate 60-80% of the privacy budget and an ε value < 0.5 (ensuring strong privacy with a 5-10% loss of accuracy);

[0064] Low-sensitivity data (such as weather information): Allocate 10-30% privacy budget and ε value 1-2 (balance accuracy and cost, accuracy loss <3%).

[0065] In the case implementation, the federated learning module includes cross-modal federated learning, incremental model update, and adaptive aggregation strategy;

[0066] Cross-modal federated learning is equipped with a modal feature converter and a heterogeneous gradient compensator. The modal feature converter is used to unify the feature space of multi-source heterogeneous data in the Internet of Vehicles (IoV). In IoV scenarios, vehicles collect images (cameras), point clouds (lidar), millimeter-wave radar data, and vehicle status data from the CAN bus. These data modalities vary greatly. Through the converter, the visual features of the image and the three-dimensional spatial features of the point cloud can be mapped to a unified feature dimension, allowing different modal data to participate in joint training.

[0067] The processed data includes original / pre-processed data such as image pixel matrix, point cloud coordinate set, radar echo sequence, etc.

[0068] The heterogeneous gradient compensator is used to compensate for training bias caused by the heterogeneity of connected vehicle devices. When training on-board terminals (with weak computing power and limited data) and roadside units (with strong computing power and abundant data), the gradient update amplitude and direction differ significantly. The compensator analyzes metadata such as device computing power and data quality to enhance or correct the gradients of weak devices, ensuring convergence of federated training.

[0069] Among them, the processed data includes device computing power parameters (such as GPU computing power and memory), local training gradients, and data quality assessment indicators (such as label accuracy and data completeness).

[0070] The incremental model update is equipped with a model difference analyzer and progressive weight clipping. The model difference analyzer is used to monitor the impact of changes in the connected vehicle environment on the model. When connected vehicles are driving, traffic scenes (such as morning and evening rush hours, rain and snow) and road topology (urban roads / highway switching) change dynamically, and data distribution continues to drift. The analyzer compares the prediction differences between the global model and the local model on the new data, identifies the trigger point of the scene change, and decides whether to initiate the model update.

[0071] The processed data includes global model parameter snapshots, local model prediction results, and real-time traffic scene labels (such as congestion level and weather type).

[0072] Progressive weight pruning is used to efficiently update models while controlling communication overhead. Given the limited communication bandwidth between connected vehicles (e.g., latency fluctuations in V2X communication), redundant weights (e.g., model parameters that are insensitive to traffic scenario changes) are pruned, and only critical gradients / parameter updates are transmitted. This reduces the amount of data transmitted while ensuring model accuracy, thus meeting real-time requirements.

[0073] The processed data includes the model weight matrix, weight importance score (such as calculated by Fisher information matrix), and communication bandwidth monitoring data.

[0074] Adaptive aggregation strategies are used to dynamically select aggregation rules based on the real-time status of the Internet of Vehicles (device online rate, network latency, and data privacy level). For example:

[0075] When vehicles are moving at high speeds (such as on highways), the device online time is short. Therefore, a "fast aggregation + local caching" strategy is adopted to prioritize model updates.

[0076] Involving user sensitive data (such as passenger behavior in the car), enhancing privacy protection (such as aggregation algorithms combined with differential privacy), and balancing model effectiveness and privacy security;

[0077] Processed data includes device online status (heartbeat packet feedback), network delay monitoring values, and data privacy tags (such as sensitive / non-sensitive data classification).

[0078] By "unifying multimodal data → correcting device differences → dynamic incremental updates → intelligent aggregation and matching", federated learning can be more adapted to the dynamic, heterogeneous, and privacy-sensitive actual scenarios of connected vehicles, supporting the efficient training and iteration of vehicle-road collaboration and autonomous driving models.

[0079] In the case implementation, the incentive and audit module includes contribution quantification model, traceable security audit, abnormal behavior detector and risk heat map;

[0080] The contribution quantification model is used to integrate the three dimensions of data volume, data quality, and computing resource consumption, and to incentivize participants through a reputation points system;

[0081] Traceable security audits use consortium chains to record the entire life cycle of model training and support regulatory nodes to verify the compliance of privacy protection measures;

[0082] The abnormal behavior detector is used to monitor anomalies in real time, "draw red lines" for the behavior of participating nodes, and use machine learning models to learn normal participation patterns (data distribution, gradient change patterns, communication frequency). Once a node uploads malicious data (fake road conditions, distorted driving behavior), attacks training (poisoning to invalidate the global model), or illegally steals privacy (reversely deduce sensitive information of other nodes), such as "the gradient of a certain vehicle terminal suddenly becomes abnormally large, which is inconsistent with the data pattern of the vehicle model", an alarm will be triggered immediately, and the risk will be eliminated at the source: before the malicious behavior affects the global model, the "problem node" will be intercepted in advance, isolated, and verified before deciding whether to re-connect, to ensure the model security and data security of federated learning.

[0083] Risk heat maps are used to visualize risk distribution. They mark the node locations (urban sections, regions), risk types (data leakage, model poisoning, device intrusion), and risk levels (high / medium / low) in the Internet of Vehicles on a GIS map. For example, "XX elevated bridge section, three vehicle-mounted terminal data are abnormal, and the risk of model poisoning is high." This intuitively presents the "hardest hit areas" of risk and enables precise resource allocation. By looking at the heat maps, automakers and traffic management departments can strengthen supervision in a targeted manner (add security checks and encryption levels to nodes in high-risk areas), optimize privacy policies (adjust the federated learning aggregation frequency / algorithm in risk-concentrated areas), and use limited resources to build a strong regional risk defense line.

[0084] By setting up four sub-modules, from "incentive participation (contribution) → process supervision (audit) → anomaly interception (detection) → global risk control (heat map)", the entire process of "participation motivation, safety traceability, and risk prevention and control" of the federated learning of connected vehicles is covered, making privacy-protected federated learning something that people are willing to play and can play safely.

[0085] In the case implementation, intelligent driving decision-making is used to make autonomous driving "smart and safe", privacy protection is integrated into the driving decision-making process, privacy protection technology is used to predict traffic flow, and multi-vehicle data is aggregated for traffic analysis to serve smart cities. V2X communication is used to ensure privacy security during vehicle-to-vehicle and vehicle-to-road (V2V, V2R) communications. Multi-dimensional situational awareness is used to enable the Internet of Vehicles to "see global risks", and privacy protection and security monitoring are both taken into account. The data lineage tracking dashboard is used to "record the family tree" of Internet of Vehicles data and trace the privacy handling status throughout the entire process.

[0086] By setting up five modules, the application scenarios of connected vehicles, from "driving decision-making" to "data flow", from "single-vehicle intelligence" to "vehicle-road collaboration", are all embedded with privacy protection logic. Simply put, every business function of the connected vehicle is operated with a "privacy protection shield", which not only realizes the efficiency of intelligent transportation, but also protects the security bottom line of user data.

[0087] In the case implementation, the road condition spatiotemporal encoder is used to "standardize and privacy encode" the data, turning the "messy road condition data" in the Internet of Vehicles into a format that can be used to train models while protecting privacy. The stream data processing engine is used to "screen data and resist interference" in real time to cope with the "tide of real-time data" in the Internet of Vehicles, quickly process, denoise, and select useful information. The edge model lightweight container is used to enable the model to "run at the edge without getting stuck", and to split the "big model" of federated learning into a lightweight version that can run on edge devices (on-board terminals, RSUs) to achieve real-time inference + local training. The federated data directory service is used to "map + manage permissions" for the data, and to perform "directory indexing + permission management" for the data participating in federated learning on the edge side, so that the data can be "found, used correctly, and not exceed authority."

[0088] By setting up four modules, the edge computing layer becomes the "frontline processing factory" for privacy protection of connected cars, such as:

[0089] First encode and process to turn "raw private data" into "safe and usable raw materials";

[0090] Lightweight models allow AI capabilities to be deployed at the edge, enabling real-time responses and saving resources.

[0091] Finally, the directory service manages the data to determine whether it can be used and how it should be used.

[0092] In this way, the requirements of "data real-time and privacy" of the Internet of Vehicles are solved, and a solid foundation is laid for upper-level federated learning and intelligent applications, allowing the entire privacy protection system to run "fast and securely".

[0093] In the case implementation, the on-board terminal group is the "data collector + edge node" of the connected car, the OBU device is the on-board unit, the "connector" for communication between the car and the outside world, and the main function is V2X communication (car to car, car to road). The RSU edge node is the roadside unit, which is the "edge brain" of the Internet of Vehicles, performing real-time data processing + local collaboration on the roadside. The cloud server is the "central brain" of the connected car, performing global model training + big data analysis. The third-party data platform is the "external aid" to supplement the data, providing related data outside the Internet of Vehicles (such as weather and map updates). The supervision node is the "inspector" of privacy protection, supervising the data flow + compliance throughout the process. The PKI certificate service center is the "ID card center" of network communication, issuing digital certificates for identity authentication. The data sensitivity label manager is used to "label and grade" data to guide privacy protection strategies.

[0094] The data and device layer is the "foundation + raw material warehouse + gatekeeper + supervisor" of the privacy protection of connected cars. The on-board terminal, OBU, and RSU are the "front-line troops", responsible for collection, communication, edge processing, and guarding the first line of privacy; the cloud and third parties are "backup + external aid", providing computing power and supplementary data, but they must be used in compliance under supervision; the supervision node, PKI center, and tag manager are "rule guardians", ensuring the legality of data flow, trustworthy communication, and accurate protection. All components work together to make the upper-level federated learning and privacy protection strategies "have data available, security guarantees, and rules to rely on" and truly take effect.

[0095] During implementation, follow these steps:

[0096] Step S1: Multi-source data acquisition:

[0097] Vehicle-mounted terminal cluster (camera, radar) collection: visual data (road condition images, pedestrian / vehicle recognition), radar point cloud data (3D road modeling, obstacle distance);

[0098] OBU / RSU equipment: V2X communication data (vehicle-to-vehicle collaboration, vehicle-road collaboration messages), traffic flow, and road status data;

[0099] Step S2: Device access and identity authentication:

[0100] PKI Certificate Service Center: distributes digital certificates to vehicle terminals, RSUs, and the cloud to verify device identities (to prevent forgery and man-in-the-middle attacks);

[0101] Data Sensitivity Label Manager: Marks data sensitivity levels (high sensitivity: face / license plate; medium sensitivity: driving trajectory; low sensitivity: weather);

[0102] Step S3: Real-time processing of streaming data:

[0103] Through the stream data processing engine: filtering invalid data (repeated collection, sensor false alarms), real-time denoising (such as radar point cloud clutter removal, image deblurring);

[0104] Step S4: Unify the data format (time stamp, geographic coordinate system alignment) through road condition spatiotemporal coding, and perform preliminary privacy processing (such as license plate blurring and trajectory desensitization);

[0105] Step S5: Run the lightweight federated learning model (such as local traffic flow prediction and obstacle recognition) through the edge model lightweight container to generate edge-side model parameters / gradients (without uploading the original data);

[0106] Step S6: Use the federated data directory service to record edge-side data / model resources (location, type, permissions), provide a "data map" for the core technology layer, and support on-demand calls;

[0107] Step S7: Perform geographic grid partitioning (dividing the collaboration domain by road topology, such as urban intersections and highway sections) and time sliding window (adjusting the collaboration period according to peak / off-peak traffic) through the dynamic domain partitioning strategy module;

[0108] Step S8: Dynamically switch the communication protocol of the heterogeneous network adaptation module (5G high bandwidth / V2X low latency / edge weak connection), and schedule network slices (allocate dedicated slices for federated learning to ensure bandwidth / latency);

[0109] Step 9: Differential privacy (DP) is implemented through the hybrid privacy technology stack of the privacy computing pool module: adding noise to raw data (such as driving behavior data), secure multi-party computation (MPC) for joint modeling (multi-vehicle collaborative training), zero-knowledge proof (ZKP) for model legitimacy (to prevent poisoning attacks), and dynamic privacy budget allocation (allocating privacy protection strength according to data sensitivity level).

[0110] Step S10: Through cross-modal alignment of the federated learning module (visual model + radar point cloud model → multi-sensor fusion), incremental model update (only gradient changes are transmitted to reduce communication overhead), and adaptive aggregation strategy (aggregation rules are selected according to device performance / network status);

[0111] Step S11: Contribution quantification (points are awarded to nodes based on data quality and computing power input), security audit (full-process data / model flow tracing and violation prevention), and abnormal behavior detection (identification of malicious nodes / poisoning attacks) are performed through the incentive and audit module.

[0112] Step S12: Cross-verify multi-device data (such as camera and radar data consistency) through a multi-source authenticity verification module to identify false data (to prevent data injection attacks);

[0113] Step S13: Outputting a privacy-protected decision model through intelligent driving decision-making, integrating multi-source perception data (vision + radar + V2X), and implementing functions such as lane keeping, obstacle avoidance, and traffic light recognition;

[0114] Step S14: Aggregate edge-side desensitized data through traffic flow prediction, predict congestion trends and provide scheduling suggestions (such as traffic light optimization) to the traffic management department;

[0115] Step S15: Through privacy-preserving V2X secure communication, anonymized V2X messages are obtained (hiding the vehicle's true identity) and sensitive information (such as emergency brake warnings) is encrypted and transmitted.

[0116] Step S16: Real-time risk monitoring is performed through multi-dimensional situational awareness, integrating vehicle, network, and environmental data to generate a risk heat map and warn of privacy leakage risks (e.g., data on a certain road section is eavesdropped;

[0117] Step S17: Conduct full-process audit and traceability through data lineage tracing, recording the entire chain from data collection to application (who used it and how it was processed), and support regulatory compliance (such as GDPR and ISO27001 audits).

[0118] In summary, this privacy-preserving distributed federated learning collaborative strategy for connected vehicles achieves real-time, multi-dimensional perception of the traffic environment by integrating technologies such as V2X communication, road condition spatiotemporal encoders, and streaming data processing engines, providing accurate data support for intelligent driving decisions. Through the state domain strategy module and spatiotemporal domain adaptive partitioning technology, it can dynamically adjust resource allocation and task scheduling according to real-time road conditions and network conditions, improving system response speed and efficiency. Through the privacy computing pool module, dynamic privacy budget allocation, PKI certificate service center and other technologies, the system achieves strict privacy protection in the process of data sharing and federated learning, ensuring compliance with regulatory requirements.

[0119] In addition, through technologies such as cross-modal federated learning and heterogeneous gradient compensators, it supports the efficient fusion and collaborative training of multi-source data (such as visual models and radar point cloud models), improves the model generalization capability, and optimizes the computing resource utilization of edge nodes through edge model lightweight containers and heterogeneous network adaptation modules, reduces latency, and is suitable for the deployment requirements of vehicle terminals and OBU / RSU equipment. Through functions such as data lineage tracking and review, contribution quantification model and abnormal behavior detector, it ensures the traceability of data flow and model training process, which is convenient for supervision and auditing.

[0120] In addition, dynamic optimization of computing and communication resources is achieved through technologies such as network slicing resource scheduling and multi-active disaster recovery scheduler, ensuring the stability of the system under high load or network fluctuations. Through cross-modal alignment layer and multi-sensor fusion reasoning, the system can comprehensively utilize heterogeneous data such as vision and radar point cloud to improve the accuracy and robustness of environmental perception, thereby enabling the system to build an intelligent driving solution covering the entire chain of "data collection-edge processing-collaborative learning-privacy protection-application decision-making", which is real-time, secure, adaptable and scalable, and suitable for complex Internet of Vehicles environments.

[0121] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0122] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A privacy-preserving distributed federated learning collaborative strategy for connected cars, including a privacy protection system for connected cars, characterized by: The privacy protection system of the connected car includes an application scenario layer, a core technology layer, an edge computing layer, and a data and device layer; The application scenario layer includes: intelligent driving decision-making, communication traffic prediction, V2X communication, multi-dimensional situational awareness and data lineage tracking dashboard; The core technology layer includes: dynamic domain strategy module, heterogeneous network adaptation module, privacy computing pool module, federated learning module, incentive and audit module and multi-active disaster recovery scheduler; The edge computing layer includes: a road condition spatiotemporal encoder, a stream data processing engine, an edge model lightweight container, and a federated data directory service; The data and device layer includes: vehicle-mounted terminal group, OBU equipment, RSU edge node, cloud server, third-party data platform, supervision node, PKI certificate service center and data sensitivity label manager.

2. The privacy-preserving distributed federated learning collaborative strategy for connected vehicles according to claim 1 is characterized by: The dynamic domain division strategy module includes adaptive division of spatiotemporal domains, data sensitivity classification and cross-domain coordination mechanism; The spatiotemporal domain adaptive partitioning is provided with a spatiotemporal indexing engine for dynamically constructing a three-level domain model of "vehicle terminal-roadside unit-cloud" based on the vehicle's driving trajectory and road topology. For example, a long chain domain is formed in a highway scenario, and a star domain is formed at a city intersection. The data sensitivity classification includes sensitive units, semi-sensitive units and public units. The sensitive units contain driving behavior data (braking frequency, steering wheel angle), the semi-sensitive units contain environmental perception data (road condition images, radar point clouds), and the public units contain traffic regulation data (traffic light status and speed limit signs). The cross-domain collaboration mechanism can realize the trusted transmission of cross-domain model parameters through blockchain smart contracts and adopt homomorphic encryption technology to ensure the security of inter-domain communication.

3. The privacy-preserving distributed federated learning collaborative strategy for connected vehicles according to claim 1, characterized in that: The heterogeneous network adaptation module includes dynamic switching of communication protocols and network slice resource scheduling; The dynamic switching of communication protocols includes 5G high-bandwidth environments, V2X low-latency environments, and weak edge network connections; The network slice resource scheduling allocates dedicated network slices for federated learning tasks and achieves precise control of bandwidth and latency through software-defined networking (SDN).

4. The privacy-preserving distributed federated learning collaborative strategy for connected vehicles according to claim 1, characterized in that: The privacy computing pool module includes a hybrid privacy technology stack and dynamic privacy budget allocation; The hybrid privacy technology stack includes a data preprocessing unit, a calculation execution unit, and a result verification unit. The data preprocessing unit is used to add noise to the original data using differential privacy (DP). The calculation execution unit is used to implement joint modeling using secure multi-party computation (MPC). The result verification unit is used to verify the legitimacy of the model using zero-knowledge proof (ZKP). The dynamic privacy budget allocation is used to dynamically adjust the privacy budget according to data sensitivity and task importance. For example, 80% of the budget is allocated to steering wheel angle data and 20% is allocated to weather data. A performance monitor and a compliance evaluator are also set up. The performance monitor is used to calculate the accuracy loss in real time, and the compliance evaluator is used to connect to the supervision node.

5. The privacy-preserving distributed federated learning collaborative strategy for connected vehicles according to claim 1, characterized in that: The federated learning module includes cross-modal federated learning, incremental model updating, and adaptive aggregation strategy; The cross-modal federated learning is provided with a modal feature converter and a heterogeneous gradient compensator. The modal feature converter is used to unify the feature space of multi-source heterogeneous data of the Internet of Vehicles. In the Internet of Vehicles scenario, vehicles will collect images (cameras), point clouds (lidar), millimeter-wave radar data, vehicle status data of the CAN bus, etc. These data modalities are very different. Through the converter, the visual features of the image, the three-dimensional spatial features of the point cloud, etc. can be mapped to a unified feature dimension, so that different modal data can participate in joint training. The heterogeneous gradient compensator is used to compensate for the training deviation caused by the heterogeneity of Internet of Vehicles devices. When the on-board terminal (weak computing power and little data) and the roadside unit (strong computing power and rich data) are trained, the gradient update amplitude and direction are very different. The compensator enhances / corrects the weak device gradient by analyzing the metadata such as the device computing power and data quality to ensure the convergence of the federated training. The incremental model update is provided with a model difference analyzer and progressive weight clipping. The model difference analyzer is used to monitor the impact of changes in the Internet of Vehicles environment on the model. When the connected car is driving, the traffic scene (such as morning and evening rush hours, rainy and snowy weather) and road topology (urban road / highway switching) change dynamically, and the data distribution continues to drift. The analyzer compares the prediction difference between the global model and the local model on the new data, identifies the scene change trigger point, and decides whether to start the model update. The progressive weight clipping is used to efficiently update the model while controlling the communication overhead. The communication bandwidth between Internet of Vehicles devices is limited (such as delay fluctuations in V2X communication). Redundant weights (such as model parameters that are insensitive to traffic scene changes) are clipped, and only key gradients / parameter updates are transmitted. While ensuring model accuracy, the amount of transmitted data is reduced to adapt to real-time requirements; The adaptive aggregation strategy is used to dynamically select aggregation rules based on the real-time status of the Internet of Vehicles (device online rate, network latency, and data privacy level).

6. The privacy-preserving distributed federated learning collaborative strategy for connected vehicles according to claim 1, characterized in that: The incentive and audit module includes a contribution quantification model, traceable security audit, abnormal behavior detector and risk heat map; The contribution quantification model is used to integrate the three dimensions of data volume, data quality, and computing resource consumption, and motivate participants through a reputation points system; The traceable security audit uses the consortium chain to record the entire life cycle of the model training, supporting regulatory nodes to verify the compliance of privacy protection measures; The abnormal behavior detector is used to monitor anomalies in real time, "draw red lines" for the behavior of participating nodes, and use machine learning models to learn normal participation patterns (data distribution, gradient change patterns, and communication frequencies). Once a node uploads malicious data (fake road conditions, distorted driving behavior), conducts attack training (poisoning the global model to invalidate it), or illegally steals privacy (reversely derives sensitive information from other nodes), for example, "the gradient of a certain vehicle-mounted terminal suddenly becomes abnormally large, inconsistent with the data patterns of the vehicle model," an alarm is immediately triggered, and the risk is eliminated at the source: before the malicious behavior affects the global model, the "problem node" is intercepted in advance, isolated, and verified before deciding whether to re-access it, thereby ensuring the model and data security of federated learning. The risk heat map is used to visualize risk distribution, marking the node locations (urban sections, regions), risk types (data leakage, model poisoning, device intrusion), and risk levels (high / medium / low) in the Internet of Vehicles on the GIS map. For example, "XX elevated bridge section, 3 vehicle-mounted terminal data are abnormal, and the risk of model poisoning is high", intuitively presenting the risk "hard-hit areas" and accurately allocating resources. By looking at the heat map, car companies and traffic management departments can strengthen supervision in a targeted manner (add security verification and encryption levels to nodes in high-risk areas), optimize privacy policies (adjust the federated learning aggregation frequency / algorithm in risk-concentrated areas), and use limited resources to build a strong regional risk defense line.

7. The privacy-preserving distributed federated learning collaborative strategy for connected vehicles according to claim 1, characterized in that: The intelligent driving decision-making is used to make autonomous driving "smart and safe", and privacy protection is integrated into the driving decision-making process. The traffic flow prediction uses privacy protection technology, aggregates multi-vehicle data for traffic analysis, and serves smart cities. The V2X communication is used to ensure privacy security during vehicle-to-vehicle and vehicle-to-road (V2V, V2R) communications. The multi-dimensional situational awareness is used to enable the Internet of Vehicles to "see global risks", focusing on both privacy protection and security monitoring. The data lineage tracking dashboard is used to "record the family tree" of Internet of Vehicles data and trace the privacy handling status throughout the entire process.

8. The privacy-preserving distributed federated learning collaborative strategy for connected vehicles according to claim 1, characterized in that: The road condition spatiotemporal encoder is used to "standardize and privacy encode" data, converting the "messy road condition data" in the Internet of Vehicles into a format that can be used to train models while protecting privacy. The stream data processing engine is used to "screen data and resist interference" in real time to cope with the "tide of real-time data" in the Internet of Vehicles, quickly process, denoise, and select useful information. The edge model lightweight container is used to enable the model to "run at the edge without getting stuck", breaking down the "big model" of federated learning into a lightweight version that can run on edge devices (on-board terminals, RSUs) to achieve real-time inference + local training. The federated data directory service is used to "map and manage permissions" for data, and perform "directory indexing and permission management" for data participating in federated learning on the edge side, so that the data can be "found, used correctly, and not exceed authority." 9. The privacy-preserving distributed federated learning collaborative strategy for connected vehicles according to claim 1, characterized in that: The on-board terminal group is the "data collector + edge node" of the connected car, the OBU device is the on-board unit, the "connector" for communication between the car and the outside world, and its main function is V2X communication (car to car, car to road). The RSU edge node is the roadside unit, which is the "edge brain" of the Internet of Vehicles, performing real-time data processing and local collaboration on the roadside. The cloud server is the "central brain" of the connected car, performing global model training and big data analysis. The third-party data platform is the "external aid" for supplementary data, providing related data outside the Internet of Vehicles (such as weather and map updates). The supervision node is the "inspector" of privacy protection, supervising data flow and compliance throughout the process. The PKI certificate service center is the "ID card center" of network communication, issuing digital certificates for identity authentication. The data sensitivity label manager is used to "label and grade" data to guide privacy protection strategies.

Images