Biorobot robust motion prediction method incorporating adversarial training

Abstract

A biorobot robust motion prediction method incorporating adversarial training. In a harsh adversarial environment with human interference, for a biorobot motion behavior prediction task in the adversarial environment, a biorobot motion image captured by a camera is used as an input, and a proxy defense model is introduced, so as to design an adversarial training algorithm capable of performing robust behavior prediction on the motion process of a biorobot.

Description

A robust motion prediction method for bio-robots integrated with adversarial training Technical Field

[0001] The present invention belongs to the fields of image processing and computer vision, and in particular relates to a bio-robot robust motion prediction method integrated with adversarial training. Background Art

[0002] With the development of machine learning and large-scale deep neural network models, various computer vision-related technologies have made great progress, such as human motion behavior prediction and target detection, and the accuracy of classification and detection has also been significantly improved. In addition, with the development of implantable brain-computer interfaces in the past two years, coarse-scale behavior control of small organisms through brain-computer interfaces has become increasingly feasible. Researchers have begun to focus on the control and prediction of the motion behavior of small biological robots, which has great potential in manipulating robots to perform rescue and exploration tasks that humans cannot participate in. With sufficient experience in processing various human motion analysis tasks, the accurate prediction and analysis of the motion behavior of smaller organisms has become a new hot issue.

[0003] At present, some technologies that use deep learning technology for motion prediction and coding have been proposed, such as the learnable universal video coding method proposed in patent CN116437102A, which can obtain the reconstruction features of video frames through efficient video coding for joint motion prediction and coding; patent CN116469173A proposes a human motion prediction system based on multi-level feature extraction, which focuses on the feature information of specific parts of the human body based on Transformer, and uses feature graph convolutional networks to effectively capture the feature information between nodes; patent CN1161178121A proposes a multi-degree-of-freedom robot motion prediction method, which can predict the current actual joint position and speed of the robot in real time, facilitating remote control of the robot.

[0004] The above inventions focus more on the human body and general robot motion prediction methods. However, small biological robots are easily disturbed by background noise because the body accounts for a relatively small proportion of the environment in the obtained motion images. Assuming that the obtained motion image input contains artificially added noise disturbances, it is difficult to detect with the naked eye, but it will greatly interfere with the results of motion prediction. Therefore, this patent considers the task of biological robot motion prediction in an adversarial environment. By designing a new adversarial training strategy, it improves the robustness of the motion prediction model itself to the disturbances present in the image input. At the same time, it designs a knowledge distillation loss function to help stabilize the adversarial training process and avoid a sharp decline in model performance due to collapse-type overfitting, thereby achieving robust behavior prediction of biological robots in adversarial environments.

[0005] Summary of the Invention

[0006] This invention aims to overcome the shortcomings of existing technologies and provide a robust motion prediction method for bio-robots that integrates adversarial training. For the task of predicting bio-robot motion behavior in harsh adversarial environments with human interference, this method uses camera-captured bio-robot motion images as input, introduces a proxy defense model, and designs an adversarial training algorithm that can robustly predict its motion process. This algorithm is a two-layer adversarial training algorithm that includes inner-layer proxy model updates and outer-layer updates. It also incorporates knowledge distillation technology to design a new adversarial training loss function to prevent overfitting during model training. This method effectively improves the robustness of the motion prediction model to adversarial perturbations of varying intensities, enhancing training stability.

[0007] The steps of the specific technical solution of the present invention are as follows:

[0008] A robust motion prediction method for biorobots integrated with adversarial training is proposed. The steps are as follows:

[0009] 1) Data Preparation

[0010] 1-1) Use the camera to shoot the i-th motion image frame of the bio-robot and record it as x i , and based on the observation results, the motion prediction category is marked, recorded as y i , get the training paired data (x i ,y i );

[0011] 1-2) Based on the labeling in step 1-1), a total of M batches of data are obtained for iterative training;

[0012] 2) Model initialization

[0013] 2-1) Initialize the motion prediction model f θ ; The structural motion prediction model f θ Using the general residual classification network PreActResNet (PARN)-18, using data x i As input, the output is the corresponding motion category f θ (x i ), use y i as a label;

[0014] 2-2) Initialize the proxy model f p =f θ , f p Structure and f θ same;

[0015] 3) Generate adversarial samples and initialize historical queues

[0016] 3-1) Initialize the history queue for storing adversarial samples and initialize the adversarial perturbation to δ0;

[0017] 3-2) Perform K cycles and execute the anti-perturbation δ k+1 The generation process, k = 0, 1, 2, ..., K, against the perturbation δ k+1 The generation process is as follows: δ k+1 =max(min(δ k+1 ,ε),-ε)

[0018] in, represents the gradient of the perturbation δ, α is the attack step size; ε is the range of the adversarial perturbation; is the loss function for adversarial attacks, function

[0019] 3-3) Generated adversarial sample δ k+1 Store in the sample history queue Q δ Each time a sample is generated, step 3-3) is executed once;

[0020] 4) Inner update proxy model

[0021] 4-1) Based on motion images and Q δ The adversarial sample dequeued from the history queue is used to calculate the distillation loss function The KL divergence function is generally used, denoted as KLDivloss(), where T is the distillation temperature and is in the form of:

[0022] 4-2) Calculate the supervised loss function based on the adversarial samples and labels removed from the queue The cross entropy function Crossentropy() is generally used in the form of:

[0023] 4-3) Aggregation Distillation Loss Function and supervised loss function Introducing the equalization coefficient μ, μ∈(0,1), the final defense loss function Expressed as:

[0024] 4-4) Use Calculate the gradient of the proxy model parameter p and update it using the optimizer. The updated parameter p′ is in the form of:

[0025] Where β is the learning rate of the proxy model parameter p, represents the gradient about the proxy model parameter p;

[0026] 5) Based on the updated proxy model parameter p' and the current prediction model parameter θ, the difference between the two is calculated and the gradient of the motion prediction model parameter θ is recorded as g θ , and initialized to 0, the parameter difference is accumulated as the gradient of the motion prediction model parameters, in the form of: g θ =g θ +θ-p′;

[0027] 6) Repeat steps 4) and 5) until the history queue is empty;

[0028] 7) The proxy model f p The parameter p′ is updated to the current motion prediction model f θ The parameter θ is of the form: p′=θ

[0029] 8) Outer layer updates the motion prediction model: based on the gradient g of the motion prediction model parameters accumulated in step 5) θ , use the optimizer to update θ to θ', which is in the form of: θ'=θ-γg θ

[0030] Among them, γ is the learning rate of the motion prediction model;

[0031] 9) After the update, return to step 1) until the number of iterations N of updating the outer layer motion prediction model is reached to obtain the bio-robot motion prediction result.

[0032] The beneficial effects of the present invention are:

[0033] This paper first introduces a proxy model and constructs a two-layer adversarial training strategy that quickly updates the inner proxy model and updates the outer motion prediction model. At the same time, it combines knowledge distillation technology to design a new aggregated adversarial defense loss function, effectively improving the robustness of the motion prediction model and reducing its sensitivity to noise in the background of the bio-robot's motion environment. The improved invention has the following features:

[0034] (1) The program is simple and easy to implement. The algorithm implementation is independent of the network structure of the motion prediction model and can be combined with different large or small motion prediction models;

[0035] (2) Low computational burden. Compared with the standard adversarial training algorithm, the introduction of extremely low computational loss can effectively improve the motion prediction model's ability to defend against adversarial perturbations of different intensities.

[0036] (3) Introducing a defense loss function designed based on the concept of knowledge distillation can effectively improve the stability of training, prevent collapse-type overfitting, and improve the prediction ability in adversarial environments;

[0037] (4) The method is highly versatile and can improve standard single-step or multi-step adversarial training algorithms and achieve consistent performance improvements. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is the overall flow chart of the adversarial training algorithm for the motion prediction model. DETAILED DESCRIPTION

[0039] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0040] The present invention proposes an adversarial training method for robust motion prediction of a bio-robot in an adversarial environment, the process of which is shown in FIG1 .

[0041] In this embodiment, the experimental data used is the motion image of the biological robot (rat) captured by the third-person top camera, which is denoted as x, and the data is annotated as y. The motion categories included in the annotation include {stationary, active, turning, moving}, and the motion prediction model for being attacked is f θ , the parameter is θ, and the proxy model is f p , the parameter is p, and the network structure used by both is the residual classification structure PreActResNet (PARN)-18, using training data x i As input, the output is the corresponding motion category f θ (x i ), use y i The type of adversarial attack used as a label is PGD, the number of attack iterations is defined as K = 10, ε is the range of the adversarial perturbation, the number of outer iterations is N = 30, the data batch used for training is M = 128, the attack step size is α = 2 / 255, and the sample history queue Q is δ The queue length is 3, the agent model f p and motion prediction model f θ The learning rates are β = 0.2 and γ = 0.8, respectively. The learning rate decay method is cyclical decay. The aggregation coefficient of the distillation loss function is μ = 0.95. The distillation temperature is T = 6.0. The optimizer used for both inner and outer layer updates is the Momentum optimizer. Based on this, a two-layer adversarial training algorithm is designed. The method includes the following steps:

[0042] 1) Experimental data preparation

[0043] 1-1) First, place the biological robot (rat) required for the experiment in the center of the playground, and use the top camera to capture its i-th motion image frame, which is recorded as x. i , and based on the observation results, the motion prediction category is marked, recorded as y i , get the training paired data (x i,y i );

[0044] 1-2) Based on the above labeling, a total of M batches of data are obtained for iterative training;

[0045] 2) Model initialization is performed only when the algorithm is executed for the first time

[0046] 2-1) Initialize the motion prediction model f θ , using the training data x i As input, the output is the corresponding motion category f θ (x i ), use y i as a label;

[0047] 2-2) Initialize the proxy model f p =f θ , using the training data x i As input, the output is the corresponding motion category f θ (x i ), use y i as a label;

[0048] 3) Generate adversarial samples and initialize historical queues

[0049] 3-1) Initialize the history queue for storing adversarial samples, and initialize the adversarial perturbation to δ0∈[-8 / 255,8 / 255];

[0050] 3-2) Perform K cycles and execute the following counter-perturbation δ k+1 The generation process, k = 0, 1, 2, ..., K, executes 3-3) once each time a sample is generated: δ k+1 =max(min(δ k+1 ,ε),-ε);

[0051] The function

[0052] 3-3) Generated adversarial sample δ k Store in the sample history queue Q δ middle;

[0053] 4) Inner update proxy model

[0054] 4-1) Based on motion images and Q δ Calculate the distillation loss function for the adversarial samples in the queue The KL divergence function is generally used, denoted as KLDivloss(), where T is the distillation temperature and is in the form of:

[0055] 4-2) Calculate the supervised loss function based on the adversarial samples and labels removed from the queue The cross entropy function Crossentropy() is generally used in the form of:

[0056] 4-3) Aggregating the two loss functions and introducing the equalization coefficient μ,μ∈(0,1), the final defense loss function can be expressed as:

[0057] 4-4) Use Calculate the gradient of the proxy model and use the optimizer to update it in the form of:

[0058] 5) Based on the updated proxy model p' and the current prediction model, the difference between the two parameters is calculated and the gradient of the motion prediction model is recorded as g θ , and initialized to 0, the difference is accumulated as the gradient of the motion prediction model in the form of g θ =g θ +θ-p′

[0059] 6) Repeat steps 4) and 5) until the history queue is empty;

[0060] 7) Update the proxy model to the parameter state of the current motion prediction model, in the form of: p′=θ

[0061] 8) Outer layer updates the motion prediction model: Based on the gradient of the motion prediction model accumulated in step 5), the optimizer is used to update the motion prediction model in the form of: θ′=θ-γg θ

[0062] 9) After updating, return to step 1) until the number of outer layer iterations is reached.

[0063] On the CIFAR10 dataset, the test results of the adversarial attack method in the embodiment after training were compared with three common existing methods using PGD-10 with ε = 8 / 255. The comparison results are shown in Table 1, where FAST-AT is the method proposed in the paper "Fast is better than free: Revisiting adversarial training", FAST-AT-GA is the method proposed in the paper "Understanding and improving fast adversarial training", and FAST-BAT is the method proposed in the paper "Revisiting and Advancing Fast Adversarial Training Through the Lens of Bi-Level Optimization". As can be seen from the table, the method of the present invention is not only more robust to adversarial attacks at the adversarial perturbation strength used during training (ε = 8 / 255), but also has a more significant improvement in robustness to attacks at a higher perturbation strength (ε = 16 / 255), indicating that the method of the present invention is more effective.

[0064] Table 1

Claims

1. A bio-robot robust motion prediction method integrating adversarial training, characterized in that: Here are the steps: 1) Data preparation 1-1) Use the camera to shoot the i-th motion image frame of the bio-robot, denoted as x i , and the motion prediction category is labeled based on the observation results, recorded as y i , get the training paired data (x i ,y i ); 1-2) Based on the labels in step 1-1), a total of M batches of data are obtained for iterative training; 2) Model initialization 2-1) Initialize the motion prediction model f θ ; The structural motion prediction model f θ Using the residual classification network PreActResNet-18, using data x i As input, the output is the corresponding motion category f θ (x i ), use y i as a label; 2-2) Initialize the proxy model f p =f θ , f p Structure and f θ same; 3) Generate adversarial samples and initialize historical queues 3-1) Initialize the history queue for storing adversarial samples and initialize the adversarial perturbation to δ0; 3-2) Perform K cycles and execute the anti-disturbance δ k+1 The generation process, k = 0, 1, 2, ..., K, against the perturbation δ k+1 The generation process is as follows: d k+1 =max(min(δ k+1 ,e),-e) in, represents the gradient of the disturbance δ, α is the step size of the attack; ε is the range of the adversarial disturbance; l atk is the loss function for adversarial attacks, function 3-3) Generated adversarial sample δ k+1 Store in the sample history queue Q δ Each time a sample is generated, step 3-3) is executed once; 4) Inner Update Proxy Model 4-1) Based on motion images and Q δ The adversarial sample dequeued from the history queue calculates the distillation loss function l kd , using the KL divergence function, denoted as KLDivloss(), T is the distillation temperature, in the form of: the kd =KLDivloss(softmax(f p (x) i +δ k ) / T),softmax(f p (x) i ) / T)) 4-2) Calculate the supervised loss function l based on the adversarial samples and labels removed from the queue sup , using the cross entropy function Crossentropy(), the form is: l sup =Crossentropy(f p (x i +δ k ),y i ) 4-3) Aggregation distillation loss function l kd And the supervision loss function l sup , introduce the equalization coefficient μ,μ∈(0,1), the final defense loss function l def It is expressed as: l def =(μ×T 2 )×l kd +(1-μ)×l sup 4-4) Use l def Calculate the gradient of the proxy model parameter p and update it using the optimizer. The updated parameter p′ is in the form of: Where β is the learning rate of the proxy model parameter p, represents the gradient of the proxy model parameter p; 5) Based on the updated proxy model parameter p' and the current prediction model parameter θ, the difference between the two is calculated, and the gradient of the motion prediction model parameter θ is denoted as g θ , and initialized to 0, the parameter difference is accumulated as the gradient of the motion prediction model parameters, in the form of: g θ g θ +θ-p′ 6) Repeat steps 4) and 5) until the history queue is empty; 7) The proxy model f p The parameter p′ is updated to the current motion prediction model f θ The parameter θ is of the form: p′=θ 8) Update the motion prediction model in the outer layer: based on the gradient g of the motion prediction model parameters accumulated in step 5) θ , use the optimizer to update θ to θ', in the form of: θ′=θ-γg θ Among them, γ is the learning rate of the motion prediction model; 9) After updating, return to step 1) until the number of iterations N of updating the outer motion prediction model is reached to obtain the bio-robot motion prediction result.

Images