A robustness quantitative analysis method for intelligent image classification system based on compressed sensing technology

Abstract

This invention discloses a robustness quantitative analysis method for intelligent image classification systems based on compressed sensing technology, belonging to the field of artificial intelligence. The method comprises the following steps: 1: collecting original image data and modifying labels; 2: training an intelligent image classification system to be tested using an original image training set and an adversarial image training set with modified labels; the intelligent image classification system comprises a transmitter, a receiver based on a compressed sensing algorithm, and an image classification end; and 3: testing the trained original intelligent image classification system and the adversarial intelligent image classification system using an original image test set, statistically calculating the misclassification rate of adversarial category images, and calculating the robustness score of the intelligent image classification system; a higher score indicates higher system robustness. The invention aims to provide guidance for the secure application of compressed sensing technology in intelligent image classification systems.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a robustness quantitative analysis method for an intelligent image classification system based on compressed sensing technology. Background Art

[0002] Intelligent image classification systems are now widely used in various fields such as autonomous driving, facial recognition, and assisted medical care. However, as data volumes continue to grow and model accuracy requirements continue to rise, the bandwidth required for image signal transmission in intelligent image classification systems is increasing, and the sampling rate and processing speed required for image information acquisition are becoming increasingly demanding. Compressed sensing technology, by exploiting the sparse nature of signals, can perfectly reconstruct signals at a sampling rate far below the Nyquist rate. Since its introduction, this technology has attracted widespread attention from both academia and industry. Applying compressed sensing technology to intelligent image classification systems is an emerging approach to addressing the challenges of image signal transmission bandwidth requirements and the sampling rate and processing speed required for image information acquisition. However, as a preprocessing step in intelligent image classification systems, compressed sensing increases the target selection of malicious attackers, reducing the robustness of intelligent image classification systems.

[0003] Therefore, obtaining the robustness quantification results of the intelligent image classification system based on compressed sensing technology has guiding significance for selecting an intelligent image classification system or selecting the optimal compressed sensing technology. Summary of the Invention

[0004] The purpose of the present invention is to address the deficiencies of the existing technology and provide a robustness quantitative analysis method for intelligent image classification systems based on compressed sensing technology. The robustness difference before and after the application of compressed sensing technology in the intelligent image classification system can be quantified into a robustness score, providing guidance for the safe application of compressed sensing technology in intelligent image classification systems.

[0005] The technical solution adopted in the present invention is:

[0006] A robustness quantitative analysis method for an intelligent image classification system based on compressed sensing technology includes the following steps:

[0007] Step 1: Collect the original image dataset and divide it into an original image training set and an original image test set; modify the labels of the original image training set, changing the image labels of one category to other misleading categories in the original category, and use the original image training set with modified labels as the adversarial image training set;

[0008] Step 2: Using the original image training set and the adversarial image training set to train the intelligent image classification system to be tested, respectively, to obtain a trained original intelligent image classification system and an adversarial intelligent image classification system; the intelligent image classification system includes a transmitter, a receiver based on a compressed sensing algorithm, and an image classification terminal;

[0009] Step 3: Use the original image test set to test the trained original intelligent image classification system and the adversarial intelligent image classification system, calculate the misclassification rate of the adversarial category images, and calculate the robustness score of the intelligent image classification system. The higher the score, the more robust the system.

[0010] Furthermore, the transmitting end in the intelligent image classification system is used to downsample the original image data through a sampling matrix to obtain a compressed image signal; the receiving end is used to receive the compressed image signal and reconstruct it based on a compressed sensing algorithm to obtain a reconstructed image signal; and the image classification end is used to classify the reconstructed image signal to complete the image classification task.

[0011] Furthermore, the compressed sensing algorithm includes a compressed sensing algorithm that relies on mathematical derivation and a compressed sensing algorithm based on deep learning; during the training process of step 2, for the compressed sensing algorithm that relies on mathematical derivation, only the sampling matrix parameters in the transmitting end are modified; for the compressed sensing algorithm based on deep learning, the sampling model parameters in the transmitting end and the compressed sensing algorithm parameters in the receiving end are modified at the same time.

[0012] Furthermore, when training the intelligent image classification system to be tested in step 2, the reconstruction loss is calculated using the reconstructed image signal output by the compressed sensing algorithm and the original image data, and the cross-entropy loss is calculated using the classification result output by the image classification end and the image label. The weighted result of the reconstruction loss and the cross-entropy loss is used as the total loss, and the system parameters in the transmitter and receiver are updated by the gradient descent method.

[0013] Furthermore, the adversarial image training set construction steps and the adversarial intelligent image classification system training steps in steps 1 and 2 are repeated. When constructing the adversarial image training set, all categories of images in the original image dataset are traversed respectively to obtain the adversarial intelligent image classification system corresponding to each category. Then, according to step 3, the adversarial category image misclassification rate under each category is calculated to calculate the robustness score of the intelligent image classification system, and the average robustness score is taken as the final score.

[0014] Furthermore, the adversarial category image misclassification rate is obtained by dividing the total number of images belonging to the adversarial category but misclassified as the misleading category by the total number of images belonging to the adversarial category.

[0015] Furthermore, the robustness score calculation formula of the intelligent image classification system is:

[0016] RobustScore=1-(E A -E O )

[0017] Among them, E O is the misclassification rate of adversarial category images of the original intelligent image classification system, E A It is the misclassification rate of adversarial category images against the intelligent image classification system.

[0018] Furthermore, the compressed sensing algorithm includes any one of OMP, ISTA, ADMM, TIP-CSNet, DeepInv and ReconNet.

[0019] Furthermore, the robustness score of the intelligent image classification system can be used to screen the optimal intelligent image classification system.

[0020] The beneficial effects of the present invention are:

[0021] In response to the current lack of methods to quantitatively evaluate the impact of compressed sensing technology on the robustness of intelligent image classification systems, this paper proposes a quantitative analysis method. This method quantifies the difference in robustness before and after the application of compressed sensing technology in intelligent image classification systems into a robustness score, providing guidance for the safe application of compressed sensing technology in intelligent image classification systems.

[0022] Figures in the specification

[0023] Figure 1 It is a structural diagram of the intelligent image classification system targeted by the present invention.

[0024] Figure 2 This is a flow chart of a robustness quantitative analysis method for an intelligent image classification system oriented to compressed sensing technology proposed in the present invention. DETAILED DESCRIPTION

[0025] The following are specific embodiments of the present invention and the accompanying drawings to further describe the technical solutions of the present invention, but the present invention is not limited to these embodiments.

[0026] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all steps. For example, some steps may be decomposed, while some steps may be combined or partially combined, so the actual execution order may change according to actual circumstances.

[0027] Figure 1This is a schematic diagram of the structure of the intelligent image classification system targeted by the present invention. The basic structure of the intelligent image classification system using compressed sensing technology consists of a transmitter, a receiver, and a downstream image classification task model based on deep learning. The transmitter downsamples the original image data using a sampling matrix to obtain a compressed image signal. The receiver receives the compressed image signal transmitted by the transmitter and reconstructs the image using a compressed sensing algorithm to obtain a restored image of the original image. Finally, the restored image is fed into the downstream image classification task model based on deep learning to complete the image classification task.

[0028] The present invention constructs an adversarial image training set for the intelligent image classification system for retraining, which is called an adversarial compressed sensing algorithm. Its target effect is to convert the pre-processed image into an adversarial sample of the deep learning image classification model in the intelligent image classification system, so that the image has the ability to induce the deep learning model to misclassify the adversarial category image as a misleading category image.

[0029] The present invention proposes a robustness quantitative analysis method for intelligent image classification systems using compressed sensing technology. The method first modifies the labels of the original dataset, changing the labels of images in the adversarial category to those in the misleading category. When training the adversarial compressed sensing algorithm, the structure of the intelligent image classification system using compressed sensing technology remains unchanged, and the transmitter and receiver of the intelligent image classification system are retrained using the image data with modified labels. The difference in robustness before and after retraining of the intelligent image classification system is quantified as a robustness score, providing guidance for the safe application of compressed sensing technology in intelligent image classification systems. A higher score indicates a higher robustness of the system using compressed sensing technology, and a lower vulnerability to compressed sensing technology.

[0030] like Figure 2 As shown, the method of the present invention comprises the following steps:

[0031] Step 1: Randomly select one category of images from the original image dataset as the adversarial category, and randomly select one category of images from the remaining image categories as the misleading category. Modify the original image dataset by changing the labels of the images in the adversarial category to the misleading category, thereby obtaining an adversarial image training set for the compressed sensing algorithm. In this embodiment, the original image dataset is divided into an original image training set and an original image test set, and the image label modification operation is performed only on the original image training set.

[0032] Step 2: Use the original image training set and the adversarial image training set to train the intelligent image classification system to be tested, respectively, to obtain the trained original intelligent image classification system and the adversarial intelligent image classification system.

[0033] In this embodiment, the structure of the original intelligent image classification system using compressed sensing technology remains unchanged, only the training set changes. The original intelligent image classification system is obtained by training with the original image training set, and the adversarial intelligent image classification system is obtained by training with the adversarial image training set. The structure of the intelligent image classification system is as follows: Figure 1 As shown, the downsampling matrix at the transmitting end and the compressed sensing reconstruction algorithm at the receiving end can adopt a variety of implementation methods in the prior art. The robust quantization proposed in the present invention is aimed at a specific intelligent image classification system, that is, the downsampling matrix and the compressed sensing reconstruction algorithm in the system are a pair of definite combinations. The present invention can obtain the optimal combination through testing.

[0034] Compressed sensing reconstruction algorithms are generally divided into three types. The first type relies on mathematical derivations, such as OMP, ISTA, and ADMM. The second type is based on deep learning, such as TIP-CSNet, DeepInv, and ReconNet. The third type mimics mathematical derivations but is actually based on deep learning, such as ISTA-Net and ADMM-Net. Different compressed sensing algorithms update different parameters during training. For the first type of compressed sensing algorithm, the parameters for parameter training are the sampling matrix at the transmitter. For the second and third types of compressed sensing algorithms, the parameters for parameter training are the sampling model at the transmitter and the compressed sensing algorithm parameters at the receiver, that is, the parameters of each neural network layer in the deep learning model. The compressed sensing algorithm obtained after training is called an adversarial compressed sensing algorithm.

[0035] The training process is repeated N times to ensure the effectiveness of the CS algorithm, where N is a parameter that can be adjusted based on the effectiveness of the CS algorithm. During each training session, the labeled data is fed into a system that allows CS algorithm parameter updates. The CS algorithm first outputs a reconstructed image, which is then classified by the deep learning-based image classification model. Finally, the optimizer uses the image reconstruction loss function calculated using the CS algorithm's reconstructed image and the cross-entropy loss function calculated using the classification results from the deep learning model in the intelligent image classification system to calculate the total model loss function.

[0036] The design philosophy behind the image reconstruction loss function is that it should reflect the difference between the image reconstructed by the compressed sensing algorithm and the images in the original dataset, as well as the smoothness of the image reconstructed by the compressed sensing algorithm. In one embodiment of the present invention, the image reconstruction loss function is composed of a weighted sum of the L1-norm distance between the reconstructed image and the original image, the value of 1 minus the structural similarity (SSIM) between the reconstructed image and the original image, and the smoothness loss function TVLoss of the reconstructed image. These three values ​​can be multiplied by a weight coefficient to change their influence on the image reconstruction loss function, and the compressed sensing algorithm parameters are updated using gradient descent.

[0037] Step 3: Use the original image test set to test the trained original intelligent image classification system and the adversarial intelligent image classification system, calculate the misclassification rate of the adversarial category images, and calculate the robustness score of the intelligent image classification system. The higher the score, the more robust the system.

[0038] In this embodiment, the misclassification rate of adversarial class images is obtained by dividing the total number of images belonging to the adversarial class but misclassified as misleading classes by the total number of images belonging to the adversarial class. The specific formula for the robustness score is:

[0039] RobustScore=1-(E A -E O )

[0040] Among them, E O is the misclassification rate of adversarial category images of the original intelligent image classification system, E A It is the misclassification rate of adversarial category images against the intelligent image classification system.

[0041] Step 4: Repeat steps 1 to 3. When constructing the adversarial image training set, traverse all categories of images in the original image dataset to obtain the adversarial intelligent image classification system corresponding to each category. Then, calculate the adversarial category image misclassification rate under each category according to step 3 to calculate the robustness score of the intelligent image classification system, and take the average robustness score as the final score.

[0042] For a fixed dataset, the more repetitions, the more the average robustness score reflects the true robustness of the current system. A higher average robustness score indicates a more robust system using compressed sensing technology and less vulnerability to compressed sensing technology.

[0043] The present invention verifies the robustness quantitative analysis method for the intelligent image classification system proposed above. Experiments show that the first intelligent image classification system based on a mathematically derived compressed sensing algorithm has the highest robustness score, the second intelligent image classification system based on a deep learning compressed sensing algorithm has the lowest robustness score, and the third intelligent image classification system based on a mathematically derived but actually deep learning compressed sensing algorithm has a robustness score between the first two. This embodiment designs a test set to test the three types of intelligent image classification systems, which is consistent with the calculated robustness scores, that is, the first type has the lowest false recognition rate, the second has the highest false recognition rate, and the third type has a false recognition rate between the first two.

[0044] The robustness quantitative analysis method of the intelligent image classification system of the present invention can be used to screen the optimal intelligent image classification system, or select the optimal combination from several sampling matrices and several compressed sensing algorithms, providing guidance for the safe application of compressed sensing technology in intelligent image classification systems.

[0045] The above examples are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above examples, and many variations are possible. All variations that can be directly derived or imagined by a person skilled in the art from the disclosure of the present invention should be considered to be within the scope of protection of the present invention.

Claims

1. A robustness quantitative analysis method for intelligent image classification systems based on compressed sensing technology, characterized by: The steps include: Step 1: Collect the original image dataset and divide it into an original image training set and an original image test set; modify the labels of the original image training set, changing the image labels of one category to other misleading categories in the original category, and use the original image training set with modified labels as the adversarial image training set; Step 2: The intelligent image classification system to be tested is trained using the original image training set and the adversarial image training set, respectively, to obtain a trained original intelligent image classification system and an adversarial intelligent image classification system; the intelligent image classification system includes a transmitter, a receiver based on a compressed sensing algorithm, and an image classification end; the transmitter in the intelligent image classification system is used to downsample the original image data to obtain a compressed image signal; the receiver is used to receive the compressed image signal and reconstruct it based on the compressed sensing algorithm to obtain a reconstructed image signal; the image classification end is used to classify the reconstructed image signal to complete the image classification task; The compressed sensing algorithm includes a compressed sensing algorithm that relies on mathematical derivation and a compressed sensing algorithm based on deep learning. During the training process, for the compressed sensing algorithm that relies on mathematical derivation, only the sampling model parameters in the transmitter are modified; for the compressed sensing algorithm based on deep learning, both the sampling model parameters in the transmitter and the compressed sensing algorithm parameters in the receiver are modified. When training the intelligent image classification system to be tested, the reconstruction loss is calculated using the reconstructed image signal output by the compressed sensing algorithm and the original image data, and the cross-entropy loss is calculated using the classification result output by the image classification end and the image label. The weighted result of the reconstruction loss and the cross-entropy loss is used as the total loss, and the system parameters in the transmitter and receiver are updated by the gradient descent method; the reconstruction loss function is composed of the weighted sum of the L1 norm distance between the reconstructed image and the original image, the value of 1 minus the structural similarity SSIM value between the reconstructed image and the original image, and the smoothness loss function TVLoss of the reconstructed image; Step 3: Use the original image test set to test the trained original intelligent image classification system and the adversarial intelligent image classification system, calculate the misclassification rate of the adversarial category images, and calculate the robustness score of the intelligent image classification system. The higher the score, the more robust the system. The robustness score calculation formula of the intelligent image classification system is: RobustScore=1-(E A -AND O ) Among them, E O is the misclassification rate of adversarial category images of the original intelligent image classification system, E A It is the misclassification rate of adversarial category images against the intelligent image classification system.

2. The robustness quantitative analysis method of the intelligent image classification system based on compressed sensing technology according to claim 1 is characterized in that: Repeat the adversarial image training set construction steps and adversarial intelligent image classification system training steps in steps 1 and 2. When constructing the adversarial image training set, traverse all categories of images in the original image dataset respectively to obtain the adversarial intelligent image classification system corresponding to each category. Then, calculate the adversarial category image misclassification rate under each category according to step 3 to calculate the robustness score of the intelligent image classification system, and take the average robustness score as the final score.

3. The robustness quantitative analysis method of an intelligent image classification system based on compressed sensing technology according to claim 1 or 2, characterized in that: The adversarial class image misclassification rate is obtained by dividing the total number of images belonging to the adversarial class but misclassified as the misleading class by the total number of images belonging to the adversarial class.

4. The robustness quantitative analysis method of the intelligent image classification system based on compressed sensing technology according to claim 1 is characterized in that: The robustness score of the intelligent image classification system can be used to screen the optimal intelligent image classification system that applies the compressed sensing algorithm.

Images