Driving vehicle road cloud cooperative safety control method and system based on trusted computing

Abstract

The invention discloses a driving vehicle road cloud collaborative safety management and control method and system based on trusted computing, and the method comprises the steps: generating an HMAC-SM3 data fingerprint through employing an SM3 Hash algorithm engine, and proposing an LSTM lightweight dynamic trust evaluation model based on a TinyML framework; through a step-by-step Hash verification and rollback mechanism, through Geohash coding and SM3 Hash joint calculation, based on an aggregation signature of a BLS12-381 curve and public key aggregation, the communication overhead during multi-RSU collaborative signature is reduced; distributed private key fragmentation decryption is realized by adding a homomorphic aggregation ciphertext; aggregating and delaying the compressed ciphertext; dynamically evaluating the credibility of the nodes by a Krum dynamic defense mechanism; the system comprises a vehicle side credible sensing layer, a roadside credible broadcast layer and a cloud credible decision-making layer. According to the invention, through layered credibility verification, dynamic trust chain extension and a data privacy protection mechanism, the security and credibility problems of multi-node cooperation in an automatic driving scene are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of the cross - application of autonomous driving and vehicle - road cooperation, and particularly to a vehicle - road - cloud cooperation security control method and system based on trusted computing. Background Art

[0002] The vehicle - road - cloud cooperation architecture is an intelligent transportation system framework that tightly combines vehicles, road infrastructure, and cloud computing resources, including the communication and cooperation among the vehicle - end, roadside, and cloud - end. The vehicle - end system perceives the surrounding environment in real - time through sensors and realizes information interaction between the vehicle and other components through the on - vehicle unit (OBU); the roadside refers to various sensing and communication devices deployed on the road, which are used for the system to obtain traffic condition information; the cloud - end is the information hub of the cooperation architecture, and makes decisions based on the information obtained by the system.

[0003] Currently, there are security risks in the information trust level in the vehicle - road - cloud architecture solution. The demand for real - time data interaction among vehicles, roadside facilities, and cloud platforms has increased sharply, but there is a lack of a real - time unified end - to - end trusted verification mechanism between vehicle - road - cloud, and there are challenges in the data trustworthiness of each end - side itself. The existing PKI (Public Key Infrastructure) system relies on a centralized certificate authority (CA), which has problems such as complex certificate management and poor cross - domain interoperability. Vehicles need to frequently apply for certificates to ensure identity trustworthiness, resulting in increased communication overhead, and static certificates are prone to leaking the vehicle's unique identifier, which is not conducive to individual privacy protection; in the vehicle - road - cloud scenario, high real - time requirements are imposed, but the existing digital signature schemes (such as ECDSA, RSA) have insufficient signature verification efficiency, and the multi - node collaborative signature mechanism in the cloud (such as PBFT consensus) has a signature generation delay of up to hundreds of milliseconds due to complex interaction processes, severely restricting the system response speed; moreover, the current vehicle - networking trust model is mostly based on static rules (such as fixed - weight scoring), lacking a hardware - level trusted root as the trust - chain anchor point, which can neither build a trusted trust - chain nor reflect the vehicle behavior credibility in real - time. Some existing patents disclose collaborative methods based on vehicle - road - cloud or vehicle - road - cloud data encryption transmission methods, but they do not solve the problem of trusted boot of hardware - layer devices, do not cover the cloud - road cooperation verification mechanism, and do not solve the problem of real - time verification at the trust level, and cannot effectively guarantee the security and privacy of data transmission, and it is difficult to resist data tampering and forgery attacks. Summary of the Invention

[0004] Objective of the Invention: Based on the above analysis, in order to construct a full-stack trusted computing architecture covering the vehicle end, roadside, and cloud end for autonomous vehicles and achieve the full-link security of data collection, transmission, and decision-making, the present invention proposes a method and system for secure control and coordination of driving vehicle-road-cloud based on trusted computing, constructs a "trinity" trusted computing architecture of vehicle end-road end-cloud end, covers the vehicle end, road end, cloud end, and trusted protocols of the three ends, and solves the security and trust issues of multi-node coordination in the autonomous driving scenario through hierarchical trusted verification, dynamic trust chain extension, and data privacy protection mechanisms; ensures the integrity of perception data of autonomous vehicles, the trustworthiness of roadside device identities, the security of cloud decisions, and the privacy of cross-domain communications.

[0005] Technical Solution: The method for secure control and coordination of driving vehicle-road-cloud based on trusted computing in the present invention includes the following steps:

[0006] 1), Adopt a sensor direct connection trusted execution environment architecture, bypass multiple protocol stacks, and perform end-to-end secure transmission through hardware-level data isolation; adopt an SM3 hash algorithm engine to generate HMAC-SM3 data fingerprints; propose an LSTM lightweight dynamic trust evaluation model based on the TinyML framework;

[0007] 1.1), Vehicle sensors access the trusted execution environment through a hardware security channel, generate a root key based on the TPM chip, derive an HMAC session key, deploy SM3 engine hardware acceleration, and generate HMAC-SM3 data fingerprints within the TEE. The data fingerprint generation method is:

[0008]

[0009] where K is the HMAC-SM3 dynamic key, ipad and opad are RFC 2104 standard padding constants, is the exclusive OR operation, and || is the data concatenation operation;

[0010] 1.2), Negotiate a new key every set time through the SM9 algorithm, append the HMAC-SM3 data fingerprint generated in step (1) to the tail of the sensor original data frame to form a secure data packet, and perform the binding of the data fingerprint and the original data:

[0011] Secure data packet = [Original data || HMAC-SM3 fingerprint]

[0012] Before loading each frame of input data, recalculate the HMAC-SM3 hash value and compare it with the fingerprint to verify the data integrity. If they are inconsistent, trigger the fuse mechanism; set the verification failure threshold. If the comparison of a single frame of data fails, discard the current frame of data and use the previous frame of valid data stored in the TEE secure memory. If the comparison of multiple consecutive frames fails, switch to the redundant sensor and cut off the contaminated data source;

[0013] The dynamic trust agent collects the CAN bus instruction frequency, the sudden increase in memory occupancy rate, and the variance characteristics of sensor data in real time as the input of the LSTM model for TinyML compression; removes the weight channels with absolute values less than 0.1 in the hidden layer of the LSTM; trains the LSTM model based on the knowledge distillation method, and uses high-temperature Softmax to soften the output distribution of the teacher model to generate soft targets:

[0014]

[0015] where q i is the probability, z is the output of the teacher model, and T is the temperature parameter;

[0016] When the vehicle safety data is transmitted externally, the data packet is encrypted in the SM4-CTR mode, and the encrypted data is sent through V2X or the CAN bus. The format is:

[0017] Transmission message = [SM4 ciphertext || TEE signature || device certificate]

[0018] Among them, the device certificate is generated based on the TPM-based AIK certificate. The TPM generates an AIK key pair, uses the EK to sign the AIK public key to generate an AIK certificate request. The privacy CA verifies the EK certificate chain and issues the AIK certificate. The receiving party verifies the AIK certificate signature through the PCA public key and synchronizes the sensor data timestamp using the PTPv2 protocol;

[0019] 2), Establish a hardware trust root based on TPM2.0, combine with the SM2 national cryptography algorithm, and ensure the trustworthiness of the RSU startup stage through hierarchical hash verification and rollback mechanism; through the combined calculation of Geohash encoding and SM3 hash, and the aggregation signature and public key aggregation based on the BLS12-381 curve, reduce the communication overhead when multiple RSUs cooperate in signing;

[0020] 2.1), After the roadside firmware RSU is powered on, TPM2.0 verifies the RSU firmware hash value level by level and compares it with the pre-stored reference hash value. If the comparison is abnormal, it rolls back to the backup partition. The comparison calculation formula is:

[0021] Verify(HASH 固件 ||HASH 配置 ) = HASH 基准

[0022] 2.2), Adopt a hierarchical signature architecture. TPM uses the SM2 private key to sign the traffic status report, and the private key is stored in the tamper-proof eFUSE memory to form the signature basic layer; convert the longitude and latitude of the RSU into a Geohash string encoding to form the signature geographical layer, fuse the basic layer and the geographical layer, and calculate the SHA-256 hash value H(m):

[0023] H(m) = SHA-256(Geohash(9) || Traffic information)

[0024] 2.3), Taking BLS12-381 as the BLS signature curve, select a random integer of order 256 on the BLS12-381 curve as the private key sk, and its corresponding public key pk is sk·G, where G is the curve base point; H(m) is the SHA-256 hash value of the geographical information generated in step, single RSU geographical signature:

[0025] σ = sk·H(m)

[0026] If multiple RSUs need to aggregate signatures for the same geographical area into one signature:

[0027]

[0028] Among them, σ agg is the BLS aggregate signature, σ i are all single RSU geographical signatures, sk i is the private key corresponding to the RUL, and the BLS aggregate signature pk agg of the public key is:

[0029]

[0030] Among them, pk i is the corresponding public key of each RUL private key sk i ;

[0031] 2.4), Broadcast the signed traffic information, in the format of [encrypted data|SHA-256 hash value|BLS aggregate geographical signature|timestamp]; The vehicle verifies the dual signature through e(G,σ agg ) = e(H(m),pk agg ), and the timestamp synchronization adopts the PTPv2 protocol to prevent replay attacks. If the verification fails, discard the data packet;

[0032] Upload the abnormal time record to the cloud, and the storage structure is:

[0033] |Abnormal event hash value|RSD ID value|Geographical Geohash|PTPv2 timestamp|

[0034] 3), Reduce the precision of the original coordinates through the Geohash algorithm, and ensure anonymity through the k-means clustering algorithm; Achieve distributed private key sharding decryption through additive homomorphic aggregation ciphertext, and compress the ciphertext aggregation delay to the millisecond level; The Krum dynamic defense mechanism dynamically evaluates the node credibility and eliminates malicious parameters;

[0035] 3.1), determine the map of the urban road network coverage area, equally divide the map into grid cells based on latitude and longitude. Each grid is assigned a Grid_ID, calculate the coordinates of the grid center point, use the Geohash algorithm to convert the center point coordinates into a 7-bit string, convert the original vehicle coordinates into a 9-bit string, intercept the first 7 characters of the 9-bit Geohash, and map the vehicle's 7-bit Geohash to the corresponding Grid_ID by comparison. The number of vehicles in each area is ≥ 50;

[0036] Before each vehicle uploads its privacy information, use Paillier homomorphic encryption:

[0037] E(c) = g c ·r n mod n 2

[0038] Among them, c is the plaintext content of the vehicle information, r is a random number, n is the public key modulus, g = 1 + n, and mod is the modulus operation; The cloud MPC receives the ciphertext data of multiple vehicles and realizes ciphertext aggregation through additive homomorphism:

[0039] E(C) = E(c1)·E(c2)···E(c k ) = E(∑c i )

[0040] Among them, E(C) is the ciphertext aggregation, E(c i ) is the ciphertext generated after Paillier homomorphic encryption of each vehicle's information, and the calculation delay is ≤ 200ms;

[0041] If decrypting, the cloud node uses distributed private key sharding for decryption:

[0042] C = L(c λ mod n 2 )·μ mod n

[0043] Among them, C is the aggregated plaintext of the vehicle, λ, μ are private key parameters, and L(x) = (x - 1) / n;

[0044] 3.2), inject Laplace noise Lap(10)(ε = 0.1) based on the differential protection mechanism:

[0045] C' = C + Lap(10)

[0046] Among them, C' is the plaintext after injecting noise;

[0047] Map the total traffic C' after decryption in each area to the geographical grid to generate a real-time heat map;

[0048] 3.3), Run the vehicle - end autonomous driving intelligent model code within the SGX Enclave; use the Krum algorithm to identify and remove malicious parameters, and calculate each parameter ΔW uploaded by the vehicle - end i The Euclidean distance from other parameters: d ij = ||ΔW i - ΔW j ||2; For each ΔW i Select the sum of the nearest n - f - 2 distances. Here, n is the total number of nodes, and f is the expected number of abnormal nodes;

[0049] 3.4), The trusted cloud issues instructions to select multi - node BLS aggregated signatures. After the vehicle - end verifies the signature consistency, it executes, and the cloud data is stored for evidence.

[0050] In step 1.1), the vehicle sensors access the trusted execution environment through a hardware security channel. The trusted execution environment verifies its own integrity through the TPM chip when the device starts, generates a root key based on the TPM chip, derives an HMAC session key, deploys the SM3 engine hardware acceleration, and generates an HMAC - SM3 data fingerprint within the TEE. The HMAC - SM3 data fingerprint is the HMAC - SM3 hash value, and the single - frame fingerprint generation delay <= 1ms.

[0051] In step 1.2), a new key is negotiated every set number of hours through the SM9 algorithm and stored in a tamper - proof secure storage area based on the electronic fuse eFUSE. The HMAC - SM3 data fingerprint generated in step 1) is appended to the tail of the sensor raw data frame to form a secure data packet.

[0052] In step 1.2), the real - time path - planning autonomous driving algorithm is executed within the TEE security area, and the code model of the real - time path - planning autonomous driving algorithm is protected by the SGX enclave.

[0053] In step 1.2), the secure data packet is stored in the secure memory of the TEE, encrypted using AES - 256 - GCM, and accessed by the authorized module.

[0054] In step 1.2), the TinyML - framework - compressed LSTM model outputs a trust result every set time, the inference time consumption <= 5ms. Neurons or connections in the LSTM neural network are removed, and weight channels with an absolute value < 0.1 in the hidden layer of the LSTM are removed, and 32 - bit floating - point parameters are converted to 8 - bit integers.

[0055] In step 3.1), determine the urban road network coverage area map and the longitude and latitude boundary values. Divide the map into 50*50 grid cells equally according to the longitude and latitude. Each grid is identified by a unique ID Grid_ID. Calculate the coordinates of the grid center point, convert the coordinates of the center point to a 7-bit string using the Geohash algorithm, convert the longitude and latitude of the original vehicle coordinates to a 9-bit string, intercept the first 7 characters of the 9-bit Geohash, complete the geographical fuzzing process of the vehicle position, and map the 7-bit Geohash of the vehicle to the 7-bit Geohash of the grid. Convert the 7-bit Geohash of the vehicle to the corresponding Grid_ID, group according to the fuzzed position, and the number of vehicles in each area ≥ 50; areas with less than 50 are merged with adjacent areas.

[0056] In step 3.3), run the vehicle-side autonomous driving intelligent model code within the SGX Enclave. Attach the TEE remote attestation SGXQuote when uploading the model parameters, and compare with the pre-stored whitelist Enclave hash; use the Krum algorithm to identify and remove potential malicious parameter poisoning attacks, and calculate each parameter ΔW uploaded by the vehicle side i The Euclidean distance from all other parameters: d ij =||ΔW i -ΔW j ||2. For each ΔW i Select the sum of the nearest n-f-2 distances (n is the total number of nodes, f is the expected number of abnormal nodes). The one with the lowest score is the most trustworthy, and the TOP5% parameters with the highest scores are removed to achieve parameter update; ΔW i Represents the parameters and geographical location of the vehicle side; among them, f is adjusted according to historical data dynamically and the vehicle-side trust score setting value. If the variance of parameters surges in multiple rounds, f is increased.

[0057] In step 3.3), run the vehicle-side autonomous driving intelligent model code within the SGX Enclave. Attach the TEE remote attestation SGXQuote when uploading the model parameters. The cloud verification uses the Intel authentication service IAS / PCK to verify the Quote, compare with the pre-stored whitelist Enclave hash, and confirm that the code has not been tampered with.

[0058] The system of the method for secure management and control of vehicle-road-cloud collaboration based on trusted computing of the present invention includes a vehicle-side trusted perception layer, a roadside trusted broadcast layer, and a cloud-side trusted decision layer;

[0059] The vehicle-side trusted perception layer includes a trusted platform module, a trusted execution environment TEE module, a sensor trusted acquisition module, a dynamic trust agent DTA module, and a secure communication gateway SCG module;

[0060] The roadside trusted broadcast layer includes a trusted platform module, a geofence signature module, a status broadcast module, distributed verification nodes, and a lightweight blockchain module;

[0061] The cloud trusted decision-making layer includes a parameter verification engine, an anomaly filtering and privacy protection module, and a trusted model management module.

[0062] Beneficial effects: Compared with the prior art, the present invention has the following advantages.

[0063] (1) The present invention realizes full-stack trusted coverage of vehicle-road-cloud based on trusted computing, completes the trinity hardware-level trusted root design of the vehicle end (TEE), roadside (TPM), and cloud end (SGX), realizes hardware-software-data extended protection, and completes the security authentication of cross-domain communication.

[0064] (2) The present invention combines a lightweight dynamic trust model (LSTM based on the TinyML framework) to construct a dynamically secure collaborative defense system, significantly improving the security of data information, the false information interception rate, and the network attack defense ability, and meeting the real-time application requirements of the autonomous driving system. Description of the Drawings

[0065] Figure 1 It is the architecture diagram of the vehicle-road-cloud collaborative security control system based on trusted computing of the present invention;

[0066] Figure 2 It is the flowchart of the vehicle-road-cloud collaborative security control method based on trusted computing of the present invention. Detailed Embodiments

[0067] The architecture of the vehicle-road-cloud collaborative security control system based on trusted computing of the present invention includes:

[0068] 1) Vehicle-end trusted perception layer: Ensure the trustworthiness and security of the entire process of vehicle local data collection, calculation, and communication, including:

[0069] The trusted platform module TPM2.0, as a hardware-level trust root, realizes the verification of secure startup and running integrity, stores keys, stores the vehicle's unique identity certificate, and supports the dynamic generation of temporary session keys.

[0070] The trusted execution environment TEE module, based on Intel SGX technology, divides the security partition of sensitive data and code in the in-vehicle computing platform, realizes physical isolation from non-secure areas such as the in-vehicle entertainment system, uses national cipher algorithms to encrypt the secure memory, the key is dynamically generated by the TPM, and integrates the SM3 national cipher algorithm hardware acceleration engine to meet the real-time requirements of the system.

[0071] The sensor trusted acquisition module is directly connected to the TEE through the hardware security module HSM, bypassing the general bus to avoid intermediate layer hijacking.

[0072] The dynamic trust agent DTA module uses a lightweight AI model of LSTM compressed by the TinyML framework (size ≤ 200KB) to monitor internal dynamic features and evaluate the security of the communication link in real time.

[0073] The secure communication gateway SCG module manages the communication links inside and outside the vehicle, uses V2X encapsulated messages, uses the AIK certificate generated based on the TPM (including the vehicle's unique ID and the current trust score) as device proof, and uses the SM9 algorithm to implement end-to-end lightweight authentication between the vehicle and other devices.

[0074] 2) Roadside trusted broadcast layer: Ensure the credibility and regional security of data acquisition, fusion, and broadcast of roadside devices such as roadside unit RSU, signal lights, and cameras, including:

[0075] The trusted platform module TPM2.0 ensures the secure startup of the roadside device firmware and stores the roadside device identity certificate (SM2 private key), geofence signature key, and dynamically generates a temporary session key.

[0076] The geofence signature module encodes the RSU coordinates into a hash string using the Geohash algorithm (precision ≤ 10 meters). The encoded geographical information (geohash encoding) is used as an input parameter for the BLS signature to ensure strong association between data and location, and uses the PTPv2 protocol (Precision Time Protocol) to synchronize timestamps.

[0077] The status broadcast module periodically publishes trusted traffic status information, and the key is dynamically generated by the TPM.

[0078] The distributed verification node cross-verifies the status of roadside devices to implement a redundant switching mechanism and data consistency verification.

[0079] The lightweight blockchain module performs data archiving and auditing.

[0080] 3) Cloud trusted decision-making layer: Achieve secure computing and sharing of multi-party data while protecting data privacy, including:

[0081] The parameter verification engine verifies the integrity and credibility of the parameters uploaded by the vehicle end based on TEE remote attestation and digital signature verification.

[0082] The anomaly filtering and privacy protection module decrypts encrypted data collaboratively by cloud nodes, uses the Paillier homomorphic encryption algorithm to generate a regional traffic flow heat map, injects Laplace noise to achieve secure multi-party computing, and uses the Krum algorithm to eliminate abnormal parameters.

[0083] The trusted model management module records operation logs through blockchain evidence storage, supporting accident tracking and compliance auditing.

[0084] The vehicle-road-cloud collaborative security control method based on trusted computing of the present invention includes the following steps:

[0085] 1) Vehicle-side trusted perception method: Adopt the sensor direct connection TEE (Trusted Execution Environment) architecture, bypass multiple protocol stacks, and achieve end-to-end secure transmission through hardware-level data isolation; Use the SM3 hash algorithm engine to generate HMAC-SM3 dynamic fingerprints, reducing the hardware resource occupancy by 15% compared with traditional methods under the same security strength; Propose an LSTM lightweight dynamic trust evaluation model based on the TinyML framework, greatly compressing the model's memory occupancy while ensuring basic performance.

[0086] 1.1) Vehicle sensors (cameras, lidar) are directly connected to the trusted execution environment (TEE) through a hardware security channel (HSM). The TEE verifies its own integrity through the TPM chip when the device starts, generates a root key based on the physical unclonable function (PUF) of the TPM chip, derives the HMAC session key, deploys the SM3 engine hardware acceleration, and generates the HMAC-SM3 data fingerprint (i.e., the HMAC-SM3 hash value) within the TEE. The single-frame fingerprint generation delay <= 1ms, and the data fingerprint generation method is:

[0087]

[0088] where K is the HMAC-SM3 dynamic key, ipad and opad are the RFC 2104 standard padding constants, ⊕ is the exclusive OR operation, and || is the data concatenation operation;

[0089] 1.2) Dynamic trust chain extension: Negotiate a new key every 1 hour through the SM9 algorithm and store it in the tamper-proof secure storage area based on the electronic fuse eFUSE. Append the HMAC-SM3 data fingerprint generated in step 1) to the tail of the sensor raw data frame to form a secure data packet, completing the binding of the data fingerprint and the raw data:

[0090] Secure data packet = [raw data || HMAC-SM3 fingerprint]

[0091] The secure data packet is stored in the secure memory of the TEE, encrypted using AES-256-GCM to prevent side-channel attacks from extracting keys or intermediate data, and only authorized modules can access it.

[0092] Execute the real-time path planning autonomous driving algorithm within the TEE security area. The algorithm code model is protected by the SGX enclave (based on Intel SGX technology). Before loading each frame of input data, the HMAC-SM3 hash value needs to be recalculated and compared with the fingerprint to verify the data integrity. If the comparison is inconsistent, the fusing mechanism will be triggered.

[0093] Safety fusing mechanism: If the comparison of a single frame of data fails, discard the current frame of data and use the previous frame of valid data stored in the TEE secure memory. If the verification comparison fails for 3 consecutive frames, switch to the redundant sensor and cut off the contaminated data source.

[0094] Record the verification results within the TEE: The statistical window is 10,000 frames (the data volume of 10 minutes), set the verification failure threshold to 0.01%, and report all failure results to the cloud.

[0095] Lightweight dynamic trust assessment model: The dynamic trust agent (DTA) collects the CAN bus instruction frequency (such as the throttle signal period), sudden increase in memory occupancy rate, and sensor data variance characteristics in real time as the input of the TinyML-compressed LSTM model. The TinyML framework-compressed LSTM model (200KB) outputs a trust score (0 - 100%) every 100ms, and the inference time consumption ≤ 5ms. By removing the neurons or connections in the LSTM neural network that have little impact on the prediction error, removing the weight channels with absolute value < 0.1 in the hidden layer of the LSTM to reduce the gating calculation amount of the LSTM model, and converting 32-bit floating-point parameters into 8-bit integers; train the LSTM model based on the knowledge distillation method, that is, use the teacher LSTM (128 units in the hidden layer) to infer the original data, and use the high-temperature Softmax to soften the output distribution of the teacher model to generate soft targets:

[0096]

[0097] where, q i is the probability, z is the output of the teacher model, and T is the temperature parameter. In the embodiment of the present invention, T = 3 is set. The student LSTM (64 units in the hidden layer) is trained through the soft target, and finally the miniaturized architecture of the autonomous driving LSTM model is completed; set the hierarchical response mechanism for the trust score output by the LSTM model. If the score is higher than 70%, it responds normally. If it is higher than 50% and lower than 70%, limit the V2X communication bandwidth. If it is lower than 50%, trigger a safe stop, and upload the score log to the cloud for storage.

[0098] Data secure transmission: When the vehicle safety data is transmitted externally, the entire data packet is encrypted in the SM4-CTR mode (the key is dynamically generated by the TPM). The encrypted data is sent through V2X or the CAN bus, and the format is:

[0099] Transmission message = [SM4 ciphertext || TEE signature || Device certificate]

[0100] Among them, the device certificate is generated based on the TPM-based AIK certificate (including the vehicle unique ID and the current trust score). The TPM generates an AIK key pair, signs the AIK public key using the EK, and generates an AIK certificate request (including the vehicle ID and the current trust score). The privacy CA (PCA) verifies the EK certificate chain and issues the AIK certificate. The certificate extension field embeds the vehicle ID (such as VIN) and the encrypted storage trust score. The receiving party verifies the AIK certificate signature through the PCA public key to confirm the certificate validity. The SM9 signature speed is 350 times per second, and the PTPv2 protocol (accuracy ≤ 1 μs) is used to synchronize the sensor data timestamp to avoid the deviation of autonomous driving operation caused by timing errors.

[0101] 2) Roadside trusted broadcast: Establish a hardware trust root based on TPM2.0, combined with the SM2 national cryptography algorithm, and ensure the trustworthiness of the RSU startup phase through hierarchical hash verification and rollback mechanism; the combined calculation of Geohash encoding and SM3 hash realizes the dual trusted insurance of preventing geographical location forgery and data tampering; the aggregated signature and public key aggregation based on the BLS12-381 curve reduce the communication overhead when multiple RSUs cooperate in signing and improve the signature verification efficiency.

[0102] 2.1) Trusted startup mechanism: After the roadside firmware RSU is powered on, the TPM2.0 verifies the RSU firmware hash value level by level and compares it with the pre-stored reference hash value to realize the trusted startup of the device. If the comparison is abnormal, it rolls back to the backup partition (time-consuming <= 2 seconds). The comparison calculation formula is:

[0103] Verify(HASH 固件 ||HASH 配置 ) = HASH 基准

[0104] 2.2) Dynamic data signature mechanism: Adopt a hierarchical signature architecture. The TPM uses the SM2 private key to sign the traffic status report (such as congestion information), and the private key is stored in the tamper-proof eFUSE memory to form the signature basic layer; convert the geographical location (latitude and longitude) of the RSU into a Geohash (9-bit) string encoding to form the signature geographical layer, splice the traffic information and the geographical encoding (that is, fuse the basic layer and the geographical layer), and calculate the SHA-256 hash value H(m):

[0105] H(m) = SHA-256(Geohash(9) || Traffic information)

[0106] 2.3) BLS Aggregate Signature: BLS12-381 is used as the BLS signature curve. A random integer of order 256 on the BLS12-381 curve is selected as the private key sk, and its corresponding public key pk is sk·G, where G is the curve base point; H(m) is the SHA-256 hash value of the geographical information generated in the previous step. The signature of a single RSU for the geographical area is:

[0107] σ = sk·H(m)

[0108] If multiple RSUs sign the same geographical area, they are aggregated into one signature:

[0109]

[0110] Among them, σ agg is the BLS aggregate signature, σ i is the signature of a single RSU for the geographical area, sk i is the private key corresponding to the RUL, and the BLS aggregate signature pk agg of the public key is:

[0111]

[0112] Among them, pk i is the corresponding public key of each RUL private key sk i ;

[0113] 2.4) Secure Broadcast and Verification: Broadcast the signed traffic information every 500 ms, in the format of [encrypted data|SHA-256 hash value|BLS aggregate geographical signature|timestamp]; Vehicles verify the dual signature through e(G,σ agg ) = e(H(m),pk agg ). The timestamp synchronization adopts the PTPv2 protocol (error ≤ 1 μs) to prevent replay attacks. If the verification fails, the data packet is discarded (the signature verification failure rate ≤ 0.1%);

[0114] The abnormal time records are uploaded to the cloud, and the storage structure is:

[0115] |Abnormal event hash value|RSD ID value|Geographical Geohash|PTPv2 timestamp|

[0116] 3) Cloud Trusted Decision: The original coordinates are reduced in precision through the Geohash algorithm to achieve location obfuscation, and the k-means clustering algorithm is used to ensure anonymity and improve the accuracy of anonymous area division; The distributed private key sharding decryption is realized by directly aggregating the ciphertext through additive homomorphy to avoid the risk of single-point leakage; The MPC technology is used to reduce the number of interaction rounds and compress the ciphertext aggregation delay to the millisecond level; The Krum dynamic defense mechanism dynamically evaluates the credibility of nodes and eliminates malicious parameters to achieve cloud anti-attack.

[0117] 3.1) Geographic Privacy Protection Mechanism: Determine the longitude and latitude boundary values of the map covering the urban road network area. Divide the map equally into 50*50 grid cells according to longitude and latitude. Each grid is identified by a unique ID (Grid_ID). Calculate the coordinates of the grid center point, use the Geohash algorithm to convert the center point coordinates into a 7-bit string, convert the original vehicle coordinates (longitude and latitude) into a 9-bit string, intercept the first 7 characters (accuracy ≈ 76 meters) of the 9-bit Geohash (accuracy ≈ 4.8 meters) to complete the geographical fuzzy processing of the vehicle position, and map the 7-bit Geohash of the vehicle to the 7-bit Geohash of the grid. Directly convert the 7-bit Geohash of the vehicle into the corresponding Grid_ID and group according to the blurred position. The number of vehicles in each area is ≥ 50, realizing anonymity with k = 50; areas with less than 50 are merged with adjacent areas.

[0118] Secure Aggregation Computing Architecture: Before each vehicle uploads privacy information, use Paillier homomorphic encryption:

[0119] E(c) = g c ·r n modn 2

[0120] Among them, c is the plaintext content of vehicle information, r is a random number, n is the public key modulus, g = 1 + n, and mod is the modulus operation; the cloud MPC (Multi-Party Secure Computation) receives the ciphertext data of multiple vehicles and realizes ciphertext aggregation through additive homomorphism:

[0121] E(C) = E(c1)·E(c2)···E(c k ) = E(Σc i )

[0122] Among them, E(C) is the ciphertext aggregation, E(c i ) is the ciphertext generated after Paillier homomorphic encryption of each vehicle's information, and the calculation delay <= 200ms;

[0123] If decryption is required, the cloud node uses distributed private key sharding for decryption:

[0124] C = L(c λ modn 2 )·μ modn

[0125] Among them, C is the vehicle aggregation plaintext, λ, μ are private key parameters, and L(x) = (x - 1) / n;

[0126] 3.2) Privacy Enhancement and Auditing: Inject Laplace noise Lap(10) (ε = 0.1) based on the differential protection mechanism to provide strong privacy protection and reduce the probability of re-identification:

[0127] C' = C + Lap(10)

[0128] Among them, C' is the plaintext after injecting noise;

[0129] Map the total traffic C' after decrypting each region to the geographical grid to generate a real-time heat map;

[0130] 3.3) Update the trusted model: The code of the vehicle-side autonomous driving intelligent model runs within the SGX Enclave. When uploading model parameters, attach the TEE remote attestation (SGX Quote). The cloud verifies the legality of the Quote and the trustworthiness of the platform using the Intel Attestation Service (IAS / PCK), compares it with the pre-stored white list Enclave hash to confirm that the code has not been tampered with; Use the Krum algorithm to identify and remove potential malicious parameter poisoning attacks, and calculate each parameter ΔW uploaded by the vehicle side i The Euclidean distance from all other parameters: d ij = ||ΔW i - ΔW j ||2. For each ΔW i Select the sum of the nearest n - f - 2 distances (n is the total number of nodes, f is the expected number of abnormal nodes). The one with the lowest score is the most trustworthy, and the TOP 5% of the parameters with the highest scores are removed to achieve parameter update; Among them, ΔW i Represents the parameters and geographical location of the vehicle side; f adjusts the set value dynamically according to historical data and the vehicle-side trust score. If the variance of multiple rounds of parameters surges, then increase f.

[0131] 3.4) Cloud instruction issuance and evidence storage: The trusted cloud selects multi-node BLS aggregate signature for issuing instructions. After the vehicle side verifies the signature consistency, it executes, and the signature verification delay <= 5ms; The cloud stores data to support judicial evidence collection.

Claims

1. A method for secure management and control of vehicle-road-cloud collaboration based on trusted computing, characterized in that: It includes the following steps: 1), Adopt a sensor direct connection trusted execution environment architecture, bypass multiple protocol stacks, and perform end-to-end secure transmission through hardware-level data isolation; use an SM3 hash algorithm engine to generate HMAC-SM3 data fingerprints; propose a lightweight dynamic trust evaluation model of LSTM based on the TinyML framework; 1.1), The vehicle sensor accesses the trusted execution environment through a hardware security channel, generates a root key based on the TPM chip, derives an HMAC session key, deploys SM3 engine hardware acceleration, and generates an HMAC-SM3 data fingerprint within the TEE. The data fingerprint generation method is: Wherein, K is the HMAC-SM3 dynamic key, ipad and opad are the padding constants in the RFC 2104 standard, is the exclusive OR operation, and || is the data concatenation operation; 1.2), Negotiate a new key through the SM9 algorithm at regular intervals, append the HMAC-SM3 data fingerprint generated in step 1) to the tail of the sensor original data frame to form a secure data packet, and bind the data fingerprint to the original data: Secure data packet = [Original data || HMAC-SM3 fingerprint] Before loading each frame of input data, recalculate the HMAC-SM3 hash value and compare it with the fingerprint. If they are inconsistent, trigger the fuse mechanism; set the verification failure threshold. If the single-frame data comparison fails, discard the current frame of data and use the previous frame of valid data stored in the TEE secure memory. If multiple consecutive frames of verification comparisons fail, switch to the redundant sensor and cut off the polluted data source; The dynamic trust agent continuously collects the CAN bus instruction frequency, sudden increase in memory occupancy, and sensor data variance characteristics as the input of the LSTM model compressed by TinyML; remove the weight channels with absolute values <0.1 in the hidden layer of LSTM; train the LSTM model based on the knowledge distillation method, and use high-temperature Softmax to soften the output distribution of the teacher model to generate soft targets; where q i is the probability, z is the output of the teacher model, and T is the temperature parameter; When the vehicle security data is transmitted externally, the packet is encrypted in the SM4-CTR mode, and the encrypted data is sent through V2X or the CAN bus. The format is: Transmission message = [SM4 ciphertext || TEE signature || Device certificate] Among them, the device certificate is generated based on the TPM's AIK certificate. The TPM generates an AIK key pair, signs the AIK public key with the EK, generates an AIK certificate request, the privacy CA verifies the EK certificate chain, issues the AIK certificate, the receiver verifies the AIK certificate signature through the PCA public key, and synchronizes the sensor data timestamp using the PTPv2 protocol; 2), Establish a hardware trust root based on TPM2.0, combine with the SM2 national cryptography algorithm, through hierarchical hash verification and rollback mechanism; through the combined calculation of Geohash encoding and SM3 hash, and the aggregation signature and public key aggregation based on the BLS12-381 curve, reduce the communication overhead when multiple RSU cooperate in signing; 2.1), After the roadside firmware RSU is powered on, the TPM2.0 verifies the RSU firmware hash value and compares it with the pre-stored reference hash value. If the comparison is abnormal, roll back to the backup partition. The comparison calculation formula is: Verify(HASH 固件 ||HASH 配置 ) = HASH 基准 2.2) Adopt a hierarchical signature architecture. The TPM uses the SM2 private key to sign the traffic report, and the private key is stored in the tamper-proof eFUSE memory, forming the signature basic layer. Convert the longitude and latitude of the RSU into a Geohash string encoding to form the signature geographical layer, and fuse the basic layer and the geographical layer to calculate the SHA-256 hash value H(m): H(m) = SHA-256(Geohash(9) || traffic information) 2.3) Select an integer of order 256 on the BLS12-381 curve as the private key sk, and the public key pk corresponding to the private key is sk·G, where G is the curve base point; H(m) is the SHA-256 hash value of the geographical information. The single RSU geographical signature: σ = sk·H(m) If multiple RSUs sign the same geographical area, they are aggregated into one signature: Among them, σ agg is the BLS aggregated signature, and σ i is the individual RSU geographical signature. sk i is the private key corresponding to the RUL, and the BLS aggregated signature pk agg of the public key is: Among them, pk i is the corresponding public key of each RUL private key sk i ; 2.4), broadcast the signed traffic information, in the format of [encrypted data | SHA-256 hash value | BLS aggregated geographical signature | timestamp]; the vehicle verifies the dual signature through e(G, σ agg ) = e(H(m), pk agg ) and adopts the PTPv2 protocol for timestamp synchronization to prevent replay attacks. If the verification fails, the data packet is discarded; The abnormal time record is uploaded to the cloud, and the storage structure is: |Abnormal event hash value|RSD ID value|Geographical Geohash|PTPv2 timestamp| 3) Reduce the accuracy of the original coordinates and ensure anonymity through the k-means clustering algorithm; decrypt the distributed private key shards through additive homomorphic aggregation ciphertext to compress the ciphertext aggregation delay; the Krum dynamic defense mechanism dynamically evaluates the node credibility and eliminates malicious parameters; 3.1) Determine the map of the urban road network coverage area, evenly divide the map into grid cells, each grid uses Grid_ID, calculate the coordinates of the grid center point, convert the center point coordinates into a 7-bit string using the Geohash algorithm, convert the original vehicle coordinates into a 9-bit string, intercept the first 7-bit characters of the 9-bit Geohash, and map the vehicle's 7-bit Geohash to the corresponding Grid_ID by comparing it with the grid's 7-bit Geohash; Before each vehicle uploads the privacy information, use Paillier homomorphic encryption: E(c) = g c ·r n mod n 2 Among them, c is the plaintext content of the vehicle information, r is a random number, n is the public key modulus, g = 1 + n, and mod is the modulus operation; the cloud MPC receives the ciphertext data of multiple vehicles and realizes ciphertext aggregation through additive homomorphism: E(C) = E(c1)·E(c2)···E(c k ) = E(Σc i ) Among them, E(C) is the ciphertext aggregation, and E(c i ) is the ciphertext generated after the vehicle is encrypted by Paillier homomorphic encryption; For decryption, the cloud node uses distributed private key shard decryption: C = L(c λ mod n 2 )·μ mod n Among them, C is the aggregated plaintext of the vehicle, λ, μ are private key parameters, and L(x) = (x - 1) / n; 3.2) Inject Laplace noise Lap(10) (ε = 0.1) based on the differential protection mechanism: C' = C + Lap(10) Among them, C' is the plaintext after injecting noise; Map the total traffic C' after decryption in each area to the geographical grid to generate a real-time heat map; 3.3) Run the vehicle - side driving intelligent model code within the SGX Enclave; use the Krum algorithm to identify and remove malicious parameters and calculate the parameter ΔW uploaded by the vehicle - side i The Euclidean distance from other parameters: d ij = ||ΔW i - ΔW j ||2; For ΔW i Select the sum of the n - f - 2 closest distances, where n is the total number of nodes and f is the expected number of abnormal nodes; 3.4) The trusted cloud issues instructions and selects multi-node BLS aggregated signatures. After the vehicle end verifies the signature consistency, it executes, and the cloud data is stored for evidence.

2. The method for driving vehicle-road-cloud collaborative security control based on trusted computing according to claim 1, wherein: In step 1.1), the vehicle sensors access the trusted execution environment through a hardware security channel. When the device starts up, the trusted execution environment verifies its own integrity through the TPM chip, generates a root key based on the TPM chip, derives an HMAC session key, deploys the SM3 engine hardware acceleration, and generates an HMAC-SM3 data fingerprint within the TEE. The HMAC-SM3 data fingerprint is the HMAC-SM3 hash value, and the single-frame fingerprint generation delay <= 1 ms.

3. The method for secure control and management of vehicle-road-cloud collaboration based on trusted computing according to claim 1, wherein: In step 1.2), a new key is negotiated every set hour through the SM9 algorithm and stored in a tamper-proof secure storage area based on the electronic fuse eFUSE. The HMAC-SM3 data fingerprint generated in step 1) is appended to the tail of the sensor raw data frame to form a secure data packet.

4. The method for secure control and management of vehicle-road-cloud collaboration based on trusted computing according to claim 1, characterized in that: In step 1.2), the real-time path planning autonomous driving algorithm is executed within the TEE security area, and the code model of the real-time path planning autonomous driving algorithm is protected by the SGX enclave.

5. The method for secure control and management of vehicle-road-cloud collaboration based on trusted computing according to claim 1, wherein: In step 1.2), the secure data packet is stored in the secure memory of the TEE, encrypted using AES-256-GCM, and provided for access by the authorized module.

6. The method for driving vehicle-road-cloud collaborative security control based on trusted computing according to claim 1, wherein: In step 1.2), the TinyML framework-compressed LSTM model outputs a trust result every set time, with the inference time <= 5 ms. Neurons or connections in the LSTM neural network are removed, and weight channels with an absolute value < 0.1 in the hidden layer of the LSTM are removed. The 32-bit floating-point parameters are converted to 8-bit integers.

7. The method for secure control and management of vehicle-road-cloud collaboration based on trusted computing according to claim 1, characterized in that: In step 3.1), the map of the urban road network coverage area and the longitude and latitude boundary values are determined. The map is evenly divided into grid cells according to the longitude and latitude. Each grid is uniquely identified by the Grid_ID. The coordinates of the grid center point are calculated, and the center point coordinates are converted into a 7-bit string using the Geohash algorithm. The longitude and latitude of the vehicle's original coordinates are converted into a 9-bit string, and the first 7 characters of the 9-bit Geohash are intercepted to complete the geographical blurring process of the vehicle's position. The 7-bit Geohash of the vehicle is mapped to the 7-bit Geohash of the grid, and the 7-bit Geohash of the vehicle is converted into the corresponding Grid_ID. The vehicles are grouped according to the blurred position, and the number of vehicles in each area >= 50.

8. The method for safe control and management of vehicle-road-cloud collaboration based on trusted computing according to claim 1, wherein: In step 3.3), the vehicle-side autonomous driving intelligent model code is run in the SGX Enclave, and the TEE remote proof SGX Quote is attached when uploading the model parameters, and the pre-stored whitelist Enclave hash is compared; the Krum algorithm is used to identify and remove malicious parameter poisoning attacks, and each parameter ΔW uploaded by the vehicle is calculated i Euclidean distance to other parameters: d ij =||ΔW i -ΔW j ||2, for each ΔW i Select the nearest nf-2 distances and sum them, where n is the total number of nodes and f is the expected number of abnormal nodes; the one with the lowest score is the most credible, and the top 5% parameters with the highest score are removed to achieve parameter update; ΔW i Represents the geographical location of the vehicle; where f adjusts the set value based on historical data dynamics and the vehicle-side trust score.

9. The method for secure control and management of vehicle-road-cloud collaboration based on trusted computing according to claim 8, wherein: In step 3.3), the vehicle-side autonomous driving intelligent model code runs within the SGX Enclave. When uploading the model parameters, the TEE remote attestation SGX Quote is appended. The cloud verifies the Quote using the Intel authentication service IAS / PCK, and compares it with the pre-stored white list Enclave hash to confirm that the code has not been tampered with.

10. A system for the method of collaborative safety control of driving vehicle-road-cloud based on trusted computing according to claim 1, characterized in that: The system includes a vehicle-side trusted perception layer, a roadside trusted broadcast layer, and a cloud-side trusted decision-making layer; The vehicle-side trusted perception layer includes a trusted platform module, a trusted execution environment TEE module, a sensor trusted acquisition module, a dynamic trust agent DTA module, and a secure communication gateway SCG module; The roadside trusted broadcast layer includes a trusted platform module, a geofence signature module, a status broadcast module, a distributed verification node, and a lightweight blockchain module; The cloud-side trusted decision-making layer includes a parameter verification engine, an anomaly filtering and privacy protection module, and a trusted model management module.

Images