A vehicle network target recognition method based on multi-export neural network security reasoning

Abstract

The present invention discloses a method for target recognition in an Internet of Vehicles (IoV) based on secure multi-export neural network reasoning. The method involves obtaining an IoV system; constructing a deep network model for the jth reasoning task of a sensor device, and constructing a system state vector for the current task j; selecting an exit for the neural network to perform the reasoning task; calculating the efficiency required by the vehicle's sensor devices to perform the current task; storing experiences containing the state vector, the neural network's exit selection strategy for the reasoning task, and the efficiency into an experience pool; and randomly sampling Z experiences from the experience pool to update the weight parameters of the deep network. This method enables the current vehicle to complete the target recognition task based on traffic data collected by the sensor device with low latency and energy consumption. The method dynamically optimizes the neural network exits for reasoning under different task categories, addressing the high overhead and low security issues associated with multi-export neural network reasoning.

Description

Technical Field

[0001] The present invention belongs to the field of network security technology, and specifically relates to a vehicle network target recognition method based on multi-export neural network security reasoning. Background Art

[0002] Currently, vehicles often use battery-powered sensor devices to collect traffic data, such as images, from the environment and process this data using deep neural networks to perform inference tasks. While deep neural networks can achieve high inference accuracy, they suffer from slow training and inference speeds. Their long inference latency makes them difficult to deploy in hardware-constrained edge devices such as real-time applications and mobile phones. Multi-output neural networks are therefore widely used to address this problem.

[0003] There is extensive research on multi-outlet neural network inference methods both domestically and internationally. Some researchers have proposed an independent dynamic early exit algorithm based on a bidirectional transformer model. This algorithm inserts an additional classification layer, known as an exit ramp, into the language model to accelerate the inference of large-scale pre-trained language models. During inference, the entropy of the output probability distribution of the current exit ramp is used to quantify the confidence of an exit in the current task's prediction. When the confidence meets the inference accuracy requirements, the inference result is returned. Others have proposed methods for converting traditional single-outlet neural networks into multi-outlet neural networks with heterogeneous and dynamic characteristics. However, the drawback of these approaches is that existing multi-outlet neural networks have high computational requirements, hindering their widespread adoption on resource-constrained mobile devices. Distributed multi-outlet neural network frameworks address this issue. For example, a distributed inference system based on existing multi-outlet neural network designs has been proposed. During inference, the client executes the first part of the model and transmits the intermediate results to a remote server. The server then continues the inference task and returns the results to the client. This algorithm allows the proportion of neural networks deployed on different devices to be adjusted based on the capabilities of the client and cloud server. However, these methods all use a single threshold for different prediction confidence distributions to make decisions. This allows attackers to launch side-channel attacks against distributed multi-exit neural networks, inferring the exit point of the current inference task and leaking model predictions. Later, researchers proposed using multiple thresholds to balance the selection of inference exit points across prediction classes and introducing confidence biases to adjust the exit probabilities of different exit points. However, these methods struggle to effectively address the security and overhead issues of multi-exit neural network inference.

[0004] To ensure that sensor devices can still maintain high inference accuracy and low overhead when using multi-export neural networks to perform inference, the present invention uses deep reinforcement learning methods to select the inference exit of the multi-export neural network for each inference task, comprehensively considering the inference efficiency, overhead and security when performing inference tasks, and achieving a balance between the three; enabling the current vehicle to complete the target recognition task of the traffic data collected by the sensor device with low latency and energy consumption. Summary of the Invention

[0005] Purpose of the invention: The present invention proposes a vehicle network target recognition method based on multi-export neural network security reasoning, which balances the reasoning efficiency, reasoning latency, reasoning energy consumption and security of the sensing system when performing reasoning tasks. The sensor equipment uses a Q network based on long-term discounted expected benefits to select the neural network outlet for reasoning tasks, thereby improving the efficiency and security of task reasoning, and thus defending against side channel attacks on multi-export neural networks.

[0006] Technical solution: The method for identifying vehicle network targets based on multi-export neural network security reasoning described in the present invention specifically includes the following steps:

[0007] (1) obtaining a vehicle networking system; the vehicle networking system includes a server and a vehicle equipped with a sensor device;

[0008] (2) Construct a deep network model, i.e., Q network, for the jth reasoning task of the sensor device, and initialize the learning rate α, discount factor δ, and the weight parameters of the Q network

[0009] (3) The sensor device constructs the system state vector for the current task j

[0010] (4) defining the exit of the neural network for reasoning tasks, and the sensor device selects the exit of the neural network for reasoning tasks;

[0011] (5) Define the channel gain from the sensor device to the server as h (k) , bandwidth is b (k) , the channel gain from the server to the sensor device is The bandwidth is

[0012] (6) If the sensor device chooses to execute the current task j on the server in step (4), the sensor device needs to set the intermediate parameters of the local neural network Send it to the server with a fixed power P; after the server completes the task, it will send the task execution result With fixed power Send to the sensor device; if the sensor device chooses to execute the current task j locally in step (4), the sensor device directly obtains the task execution result

[0013] (7) The attacker randomly infers the current task j from A neural network export outputs the task execution result;

[0014] (8) Calculate the benefits required for the vehicle's sensor equipment to perform the current task j

[0015] (9) The sensor device stores the experience including the state vector, the exit selection strategy and the benefit of the neural network in the experience pool;

[0016] (10) The sensor device randomly samples Z experiences from the experience pool to construct batch sample data;

[0017] (11) Update the weight parameters of the deep network

[0018] (12) Repeat steps (3) to (11), and add 1 to k each time until it is equal to K, so that the current vehicle can complete the target recognition task of the traffic data collected by the sensor equipment with low latency and energy consumption.

[0019] Furthermore, in step (1), the vehicle uses its own sensor equipment to collect M traffic data from the surrounding environment, processes each data, and performs the target recognition task through a ResNet18 multi-export neural network with J exits, which is deployed on the server and the sensor equipment of the vehicle respectively.

[0020] Furthermore, step (3) constructs the state vector The specific process is:

[0021] At time k, the sensor device observes the category of the original task According to the inference energy consumption of the previous moment and delay The reasoning effect of history and the probability of the attacker successfully inferring The sensor device constructs its state vector for the current task j as follows:

[0022]

[0023] Furthermore, the implementation process of step (4) is as follows:

[0024] The sensor device will transform the state vector Input into the Q network, the Q network outputs the long-term discounted expected benefits of all neural network exports in the current state And use the ε-greedy method to select an action based on this value; let the action selected by the sensor device be x j (k) , choose Q with probability 1-ε j The action with the largest value randomly selects other actions with a probability of ε, where ε∈(0,1); ε determines the exploration of the sensor device. The larger its value, the greater the randomness of the sensor device in selecting actions.

[0025] The output of the neural network for reasoning tasks is defined as in Indicates that the sensor device only executes the current task j locally, Indicates that the sensor device is executing the current task j on the server; Corresponding to the J exits of the neural network for reasoning tasks.

[0026] Furthermore, the implementation process of step (8) is as follows:

[0027] Sensor equipment uses task execution results and the actual category of the current task j Compare and obtain the current inference effect is an indicator function. When the variable in the brackets is true, the value of the indicator function is 1, otherwise it is 0. It also calculates the time delay required for the sensor equipment equipped by the vehicle to perform the current task. and energy consumption Sensor devices export tasks based on attacker inferences And the actual exit of the neural network for reasoning task j Estimate the probability of the attacker's inference being successful Then calculate the benefit at the current moment as follows:

[0028]

[0029] Among them, m1, m2, and m3 are weight coefficients, which respectively measure the importance of the inference accuracy, latency, and probability of successful inference by the attacker in the benefits of executing the current task.

[0030] Furthermore, the implementation process of step (9) is as follows:

[0031] The sensor device converts the state vector Exit selection strategy for neural network reasoning tasks benefit Constructed as an experience sequence And stored in the experience pool D. At this time, there are k experience sequences in the experience pool D.

[0032] Furthermore, the implementation process of step (10) is as follows:

[0033] The sensor device randomly samples Z pieces of experience from the experience pool D, including the historical neural network's exit selection strategy for reasoning tasks and the corresponding overhead required by the sensor device, to construct a batch sample:

[0034]

[0035] in, It represents the g(η)th experience taken from the experience pool D, and obeys the uniform distribution U(1,k), that is, g(η)~U(1,k).

[0036] Furthermore, the implementation process of step (11) is as follows:

[0037] The sensor device uses the Adam optimization algorithm to update the weight parameters of the Q network as follows:

[0038]

[0039] Among them, θ is the weight parameter and δ is the discount factor for updating the weight parameter.

[0040] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention constructs the system state by observing the category of the original task, the inference energy consumption and delay at the previous moment, the historical inference effect, and the probability of successful inference by the attacker. It combines the Q network based on the long-term discounted expected benefit value to optimize the neural network's exit selection strategy for the inference task, thereby improving the efficiency and security of task inference, and further defending against side channel attacks on multi-exit neural networks; enabling the current vehicle to complete the target recognition task of the traffic data collected by the sensor equipment with low latency and energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described in detail below with reference to the accompanying drawings.

[0044] like Figure 1 As shown, the present invention proposes a vehicle network target recognition method based on multi-export neural network security reasoning, which includes the following steps:

[0045] Step 1: Obtain an Internet of Vehicles (IoV) system. Assume that the IoV system includes a server and a sensor-equipped vehicle. The vehicle uses its own sensor equipment to collect M traffic data from the surrounding environment, such as images and videos, and processes each data point. The target recognition task is performed through a ResNet18 multi-output neural network with J outputs. The network is deployed on the server and the vehicle's sensor equipment respectively.

[0046] Step 2: The sensor device constructs a deep neural network, Q network, for the current task j, where 1≤j≤M, and initializes the learning rate α, discount factor δ, and weight parameters of the Q network And set the number of learning rounds The environment is reset with each learning round. There are K moments in a learning round, and at each moment k, the sensor device selects a neural network for the inference task for the current task j.

[0047] Step 3: Taking the jth task as an example, the sensor device constructs the current system state vector for the current task j

[0048] The method for the sensor device to construct the system state vector for the current task j can be: at time k, the sensor device observes the category of the original task According to the inference energy consumption of the previous moment and delay The reasoning effect of history and the probability of the attacker successfully inferring The sensor device constructs its state vector for the current task j as follows:

[0049]

[0050] Step 4: Define the exit of the neural network for reasoning tasks, and the sensor device selects the exit of the neural network for reasoning tasks.

[0051] The sensor device converts the state vector Input into the Q network. Among them, the Q network outputs the long-term discounted expected benefits of all neural network exports in the current state And use the ε-greedy method to select an action based on this value. Let the action selected by the sensor device at this time be x j (k) , the specific method is: select Q with a probability of 1-ε j The action with the largest value randomly selects other actions with a probability of ε, where ε∈(0,1). ε determines the exploratory nature of the sensor device. The larger its value, the greater the randomness of the sensor device's action selection. This value is often set to a small positive number.

[0052] The output of the neural network for reasoning tasks is defined as in Indicates that the sensor device only executes the current task j locally, Indicates that the sensor device is executing the current task j on the server. Corresponding to the J exits of the neural network for reasoning tasks.

[0053] Step 5: Define the channel gain from the sensor device to the server as h (k) , bandwidth is b (k) , the channel gain from the server to the sensor device is The bandwidth is

[0054] Step 6: If the sensor device chooses to execute the current task j on the server in step 4, the sensor device needs to transfer the intermediate parameters of the local neural network The server sends the task result to the server with a fixed power P. With fixed power Send to the sensor device. If the sensor device chooses to execute the current task j locally in step 4, the sensor device directly obtains the task execution result

[0055] Step 7: The attacker randomly infers the current task j from the The neural network exports the task execution results.

[0056] Step 8: The sensor device calculates the benefit for the current inference task j.

[0057] The sensor device uses the results of the previous step 7 task execution and the actual category of the current task j Compare and obtain the current inference effect is an indicator function. When the variable in the brackets is true, the value of the indicator function is 1, otherwise it is 0. The delay of executing the current task is calculated based on the parameters in steps 5 and 6. and energy consumption The sensor device exits based on the task inferred by the attacker in step 7 And the actual exit of the neural network for reasoning task j Estimate the probability of the attacker's inference being successful Then calculate the benefit at the current moment as follows:

[0058]

[0059] Among them, m1, m2, and m3 are weight coefficients, which respectively measure the importance of the inference accuracy, latency, and probability of successful inference by the attacker in the benefits of executing the current task.

[0060] Step 9: The sensor device stores the experience including the state vector, the exit selection strategy of the neural network for reasoning tasks, and the benefits into the experience pool.

[0061] The sensor device converts the state vector Exit selection strategy for neural network reasoning tasks benefit Constructed as an experience sequence And stored in the experience pool D.

[0062] Step 10: The sensor device randomly samples Z experiences from the experience pool to construct a batch sample;

[0063] The sensor device randomly samples Z experiences and constructs batch samples Among them, g(η)~U(1,k), U(1,k) is uniformly distributed.

[0064] Step 11: Update the weight parameters of the Q network using the Adam optimization algorithm or the stochastic gradient descent algorithm

[0065]

[0066] The sensor device uses the Adam optimization algorithm to update the weight parameters of the Q network as follows:

[0067]

[0068] Among them, θ is the weight parameter and δ is the discount factor for updating the weight parameter.

[0069] Step 12: Repeat steps 3 to 11, increasing k by 1 each time until it is equal to K, so that the current vehicle can complete the target recognition task of the traffic data collected by the sensor equipment with low latency and energy consumption.

[0070] In this embodiment, the Internet of Vehicles system includes a sensor device with 10,000 target recognition tasks and a server, and performs the target recognition task through a ResNet18 neural network with 4 outputs, where the first two outputs of the ResNet18 neural network are deployed on the local sensor device, and the last two outputs are deployed on the server.

[0071] Set the set of exit selection strategies for the neural network to perform reasoning tasks. For the jth reasoning task, the optional strategies of the sensor device are A j ={{0,1},{0,2},{1,3},{1,4}}.

[0072] The sensor device constructs a deep network, the Q network. The Q network consists of a fully connected layer, including an input layer, a hidden layer, and an output layer. The input layer consists of 4 neurons, the hidden layer consists of 32 neurons, and the output layer consists of 4 neurons. The learning rate α is set to 0.0001, the discount factor δ is set to 0.5, and the benefit function weight parameter m is set to 0. i1 ,m i2 ,m i3 They are 1.5, 1 and 1 respectively, and the random sampling experience number Z = 64.

[0073] Initialize the weight parameters of the Q network

[0074] The category of the current task observed by the sensor device and its corresponding channel gain and bandwidth The sensor device constructs its state vector for the current task j as follows:

[0075]

[0076] The sensor device sends the state vector Input the Q network and use the ε-greedy method to select the exit of the neural network for reasoning tasks The attacker randomly infers the current task j from the The neural network exports the task execution results.

[0077] Sensor devices utilize the results of previous task execution and the actual category of the current task j Compare and obtain the current inference effect And calculate the delay of executing the current task Energy consumption and the probability of the attacker successfully inferring Then calculate the benefit at the current moment as follows

[0078]

[0079] Sensor device construction experience sequence And store it in the experience pool D. The sensor device randomly samples 32 experiences from the experience pool to construct a batch sample

[0080] The sensor device uses the Adam optimization algorithm to update the weight parameters of the Q network. The above process is repeated, with k increasing by 1 each time until it reaches 400. This allows the current vehicle to complete the target recognition task based on the traffic data collected by the sensor device with low latency and energy consumption.

[0081] The above embodiments are only preferred embodiments of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent of the present invention.

[0082] The technical solutions provided by the embodiments of the present invention are described in detail above. Specific examples are used in this patent to illustrate the principles and implementation methods of the embodiments of the present invention. The above embodiments and features in the embodiments can be combined with each other unless there is a conflict. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A vehicle network target recognition method based on multi-export neural network security reasoning, characterized in that: The following steps are involved: (1) obtaining a vehicle networking system; the vehicle networking system includes a server and a vehicle equipped with a sensor device; (2) Construct a deep network model, i.e., Q network, for the jth reasoning task of the sensor device, and initialize the learning rate α, discount factor η, and the weight parameters of the Q network (3) The sensor device constructs the system state vector for the current task j (4) defining the exit of the neural network for reasoning tasks, and the sensor device selects the exit of the neural network for reasoning tasks; (5) Define the channel gain from the sensor device to the server as h (k) , bandwidth is b (k) ; The channel gain from the server to the sensor device is The bandwidth is (6) If the sensor device chooses to execute the current task j on the server in step (4), the sensor device needs to set the intermediate parameters of the local neural network Send it to the server with a fixed power P; after the server completes the task, it will send the task execution result With fixed power Send to the sensor device; if the sensor device chooses to execute the current task j locally in step (4), the sensor device directly obtains the task execution result (7) The attacker randomly infers the current task j from A neural network export outputs the task execution result; (8) Calculate the benefits required for the vehicle's sensor equipment to perform the current task j (9) The sensor device stores the experience including the state vector, the exit selection strategy and the benefit of the neural network in the experience pool; (10) The sensor device randomly samples Z experiences from the experience pool to construct batch sample data; (11) Update the weight parameters of the deep network (12) Repeat steps (3) to (11), and add 1 to k each time until it is equal to K, so that the current vehicle can complete the target recognition task of the traffic data collected by the sensor equipment with low latency and energy consumption.

2. The method for vehicle network target recognition based on multi-export neural network security reasoning according to claim 1 is characterized in that: In step (1), the vehicle uses its own sensor equipment to collect M traffic data from the surrounding environment, processes each data, and performs the target recognition task through a ResNet18 multi-export neural network with J exits, which is deployed on the server and the sensor equipment of the vehicle respectively.

3. The method for vehicle network target recognition based on multi-export neural network security reasoning according to claim 1 is characterized in that: Step (3) constructs the state vector The specific process is: At time k, the sensor device observes the category of the original task According to the inference energy consumption of the previous moment and delay The reasoning effect of history and the probability of the attacker successfully inferring The sensor device constructs its state vector for the current task j as follows:

4. The method for identifying vehicle network targets based on multi-export neural network security reasoning according to claim 1 is characterized in that: The implementation process of step (4) is as follows: The sensor device will transform the state vector Input into the Q network, the Q network outputs the long-term discounted expected benefits of all neural network exports in the current state And use the ò-greedy method to select an action based on this value; let the action selected by the sensor device be x j (k) , choose Q with probability 1-ò j The action with the largest value randomly selects other actions with a probability of ò, where ò∈(0,1); ò determines the exploration of the sensor device. The larger its value, the greater the randomness of the sensor device in selecting actions. The output of the neural network for reasoning tasks is defined as in Indicates that the sensor device only executes the current task j locally, Indicates that the sensor device is executing the current task j on the server; Corresponding to the J exits of the neural network for reasoning tasks.

5. The method for vehicle network target recognition based on multi-export neural network security reasoning according to claim 1 is characterized in that: The implementation process of step (8) is as follows: Sensor equipment uses task execution results and the actual category of the current task j Compare and obtain the current inference effect I(·) is an indicator function. When the variable in the brackets is true, the value of the indicator function is 1, otherwise it is 0. It also calculates the time delay required for the sensor equipment equipped on the vehicle to perform the current task. and energy consumption Sensor devices export tasks based on attacker inferences And the actual exit of the neural network for reasoning task j Estimate the probability of the attacker's inference being successful Then calculate the benefit at the current moment as follows: Among them, m1, m2, and m3 are weight coefficients, which respectively measure the importance of the inference accuracy, latency, and probability of successful inference by the attacker in the benefits of executing the current task.

6. The method for vehicle network target recognition based on multi-export neural network security reasoning according to claim 1 is characterized in that: The implementation process of step (9) is as follows: The sensor device converts the state vector Exit selection strategy for neural network reasoning tasks benefit Constructed as an experience sequence And stored in the experience pool D. At this time, there are k experience sequences in the experience pool D.

7. The method for identifying vehicle network targets based on multi-export neural network security reasoning according to claim 1 is characterized in that: The implementation process of step (10) is as follows: The sensor device randomly samples Z pieces of experience from the experience pool D, including the historical neural network's exit selection strategy for reasoning tasks and the information required by the sensor device, to construct a batch sample: in, It represents the g(η)th experience taken from the experience pool D, and obeys the uniform distribution U(1,k), that is, g(η)~U(1,k).

8. The method for identifying vehicle network targets based on multi-export neural network security reasoning according to claim 1 is characterized in that: The implementation process of step (11) is as follows: The sensor device uses the Adam optimization algorithm to update the weight parameters of the Q network as follows: Among them, θ is the weight parameter and δ is the discount factor for updating the weight parameter.

Images