Data security system based on AI and biological characteristics

Abstract

The invention discloses an identity verification method based on touch sliding gesture features in combination with AI and biological feature recognition technologies. According to the technology, multi-dimensional features of user touch behaviors are extracted through feature engineering, a full-connection neural network is innovatively adopted to deeply mine a data relationship, and a manifold learning theory and multiple dimension reduction algorithms are applied to optimize a data processing flow. In addition, a variational auto-encoder (VAE) is introduced to generate non-owner user data, so that the classification precision and generalization ability of the model are effectively enhanced. Compared with a traditional biological recognition technology (such as fingerprints, faces and irises) and behavior biological feature recognition (such as gaits and keyboard tapping rhythms), the method does not need additional hardware, is higher in privacy protection and is suitable for existing intelligent equipment. Through deep algorithm research and software testing, the system obtains a certain achievement, and an innovative scheme is provided for identity safety verification of intelligent equipment.

Description

Technical Field

[0001] This application involves the fields of artificial intelligence, biometric verification and data security, especially data security systems based on AI and biometrics. Background Art

[0002] Today, the mainstream unlocking methods are fingerprint and facial recognition, both of which rely on biometrics. Once fingerprint / facial data, which can be used as biometrics, is stolen, users face a lifelong security vulnerability—the biometric cannot be reset like a password. Dark web transaction data from 2023 shows that the price of a single valid fingerprint has fallen to $15, demonstrating that biometrics are far more reproducible than generally believed. Even more dangerously, attackers can use 3D-printed fingerprint films (which can crack most optical fingerprints with an accuracy of 5 microns) to achieve "identity hijacking," while manufacturers have yet to implement biometric invalidation mechanisms. Furthermore, beyond acquiring biometrics, traditional unlocking methods still pose certain security risks. Germany's Chaos Computer Club has demonstrated the use of wax molds and conductive ink to defeat capacitive fingerprints, while OpenAI's Sora model can generate dynamic facial data that spoofs 3D structured light. Current security authentication systems are still stuck at the stage of preventing "physical cloning" and are completely vulnerable to "digital biometrics" created by generative AI.

[0003] In view of this, how to improve the accuracy and security of the verification method has become an urgent problem to be solved. Summary of the Invention

[0004] The main technical problem solved by this application is to provide a data security system that can significantly improve the accuracy of smart device verification and ensure the real-time security of data.

[0005] In order to solve the above technical problems, the technical solution adopted in this application is: to provide a data security system, which uses an Android system-specific algorithm to sample screen touch events at intervals of 0.008 seconds, collect raw data such as the coordinate position, touch area and timestamp of each action, and generate a CSV file to the specified address A. The file name is named after the number of actions and the file sequence number, and it is cycled once every 300 items to avoid duplication. In order to maintain storage stability, a sliding window data deletion algorithm is adopted to clean up used data in a timely manner in units of 100 files to ensure that the occupied space does not exceed 500kb. Subsequently, the feature engineering algorithm reads the raw data from address A, extracts 20-dimensional high-dimensional feature data representing real touch features through secondary mathematical operations, and generates a CSV file in units of single actions and outputs it to address B. The data cleaning algorithm regularly clears invalid single-click data under address B to improve the subsequent neural network learning effect. In order to improve recognition accuracy, the project uses the VAE algorithm to generate approximate non-owner data based on the owner user data and outputs it to address C. Then, a merging algorithm combines the data from addresses B and C and stores them in address D, providing a complete dataset for the FCNN fully connected neural network. Finally, the FCNN uses supervised learning algorithms like backpropagation to extract low-dimensional biometric patterns from high-dimensional data, effectively distinguishing between the owner and non-owner data, and accurately identifying users.

[0006] The beneficial effects of this application are: extremely high resistance to simulation and cracking: this technology adopts multi-dimensional judgment and dynamic time series multi-point sampling measures, making it extremely difficult to replicate the original data. Real-time and dynamic adaptability: the system completes real-time identity authentication in a very short time by calling the real-time data of the device sensors and combining the fast matching capabilities of the deep learning model. At the same time, the model supports dynamic learning and can be continuously updated according to changes in user behavior habits to improve reliability for long-term use. It also has extremely strong privacy protection: the characteristic of behavioral data is that it is difficult to reversely restore it into personal sensitive information, and therefore it is more in line with privacy protection requirements than static biometrics such as fingerprints and faces. Even if the data is intercepted, its security risk is significantly reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments. in: Figure 1 This is the feature engineering algorithm flow chart Figure 2 It is a VAE (Variational Autoencoder) anti-data trainer based on machine master data Figure 3 It is a VAE (Variational Autoencoder) inverse data generator based on the machine master data Figure 4It is a trainer based on the FCNN fully connected neural network architecture Figure 5 It is a trainer based on the FCNN fully connected neural network architecture Figure 6 This is the overall structure diagram of the application system DETAILED DESCRIPTION

[0008] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present application.

[0009] See also Figure 1 , Figure 1 This is a flowchart for the feature engineering algorithm. The algorithm reads the raw data CSV file and calculates and generates the following 18 dimensions of feature data, reflecting the different physiological, psychological, and neural control differences of users during a single touch action. Each dimension reveals subtle differences in individual operating behaviors, which are closely related to the individual's emotional state, nervous system sensitivity, attention span, muscle coordination, and psychological state, providing essential support for the discovery and construction of low-dimensional manifolds. Finally, a secondary data CSV file containing 20 feature data (18 + 2 directional distribution ratios) is generated and output, based on a single action.

[0010] Click Frequency Mathematical formula: in, is an indicator function. If the action is a single click, the value is 1, otherwise it is 0. N is the number of the most recent 50 actions. Biometric information: This reflects the user's finger control habits and precision. Frequent tapping indicates a user's impatience, anxiety, or high finger dexterity. Psychologically, high tapping frequency is often associated with short attention spans, anxiety, or task intensity. Individual differences in mental concentration and tactile sensitivity may also influence tapping frequency.

[0011] Area Variance per Distance Mathematical symbols represent: Biometric information: Reflects the force and stability of a user's touch. Large regional variance indicates uneven finger pressure or unstable operation, which is often related to differences in muscle control, finger flexibility, or finesse. For example, some people may exhibit large pressure variations due to poor finger joint flexibility or subtle differences in the nervous system.

[0012] Area Sum per Distance Mathematical symbols represent: Biometric information: This value reflects the overall pressure applied by the user during a touch. A higher value indicates a stronger sense of touch, which may manifest as greater muscle strength or more flexible joints. It is also related to the responsiveness of the nervous system and skin sensitivity, and may manifest as comfort or fatigue when subjected to high pressure.

[0013] Distance Mathematical symbols represent: Biometric information: Reflects the range of motion during operation. Longer sliding distances are associated with differences in the user's finger flexibility and sensitivity. People with greater flexibility and range of joint motion may exhibit longer sliding paths during operation, reflecting differences in muscle coordination or joint flexibility.

[0014] Average Speed Mathematical symbols represent: Biometric information: This reflects the user's overall scrolling speed. Faster scrolling speeds are associated with higher efficiency in task execution and are associated with faster nervous system response times (such as attention and decision-making speed). An individual's mental state, level of concentration, and even physical fatigue can also influence their speed.

[0015] Maximum Speed Mathematical symbols represent: Biometric information: This reflects an individual's explosive speed during manipulation, and is generally related to their neuromuscular response, coordination, and fine control abilities. For example, some people have faster nerve reactions and can perform faster finger movements in an instant, which is related to genetic factors or motor skill training.

[0016] Acceleration Mathematical symbols represent: Among them, vi is the velocity of point ii. Biometric information: This value reflects the rate of change in velocity during an operation. Higher acceleration indicates that the user's movements are abrupt, reflecting a higher nervous system response or emotional tension (for example, anxious individuals may exhibit higher acceleration). Furthermore, differences in finger dexterity and muscle coordination may also contribute to differences in acceleration.

[0017] Fractal Dimension Mathematical symbols represent: Where LL is the trajectory length and BB is the boundary size of the trajectory. Biometric information: Reflecting the complexity of finger trajectories, higher fractal dimensions indicate more irregular movements, often associated with mood swings, attention deficit, or anxiety. Differences in nervous system sensitivity and finger movement coordination can also lead to variations in fractal dimension.

[0018] Tremor Index Mathematical symbols represent: Where θi is the angle change between adjacent points. Biometric information: It reflects the degree of finger shaking during operation. A higher shaking index is associated with physiological characteristics such as nervous system instability, anxiety or physiological tremor (such as hand tremor in the elderly).

[0019] Inertia Ratio Mathematical symbols represent: Biometric information: Reflects directional stability during sliding. A lower inertia ratio indicates more stable directional changes, which is generally related to an individual's motor control ability, concentration, and psychological stability. Individual psychological state and motor coordination ability can lead to differences in operational stability.

[0020] Touch Spread Mathematical symbols represent: Biometric information: Reflects the degree of discreteness of the touch trajectory. A larger touch extension usually indicates unstable individual operation and is associated with poor finger control accuracy, anxiety, psychological tension, or poor finger dexterity.

[0021] Time Normalized Complexity Mathematical symbols represent: Biometric information: Reflecting the relationship between trajectory complexity and time, higher complexity usually indicates that the individual's decision-making process is more complicated and is associated with higher cognitive load, level of mental activity or tension.

[0022] Path Symmetry Mathematical symbols represent: Biometric information: Reflects the symmetry of the path. Higher symmetry indicates that the user has better control when sliding, which is usually related to finger precision, emotional stability, and mental balance.

[0023] Angle Variance Mathematical symbols represent: angle_var=Var(θ i ) for i=1 to N-2 Where θi is the angular change between three adjacent points. Biometric Information: Angular variance reflects the degree of variability in the sliding trajectory. Larger angular variations often indicate unstable finger manipulation, which is related to differences in finger dexterity, joint flexibility, and differences in cerebral cortex control of movement. Smaller angular variance may indicate smoother and more consistent user manipulation, and is often associated with individuals with higher fine motor skills, such as those with better finger dexterity.

[0024] Jerk Mathematical symbols represent: Where ai is the acceleration and ti is the timestamp, indicating the rate of change of acceleration. Biometric Information: Jerk indicates the suddenness of speed changes during an operation. Higher jerk indicates abrupt or jerky user actions, which is often related to the user's emotional state (such as anxiety or tension) or the speed of their nervous system's response (such as the delicacy of motor control). Individual differences in finger dexterity, reaction speed, and physical coordination also affect jerk.

[0025] Touch Duration Mathematical symbols represent: touch_duration=max(t)-min(t) Where t is the timestamp and the duration of the touch action is calculated. Biometric Information: Touch duration reflects the user's level of focus during an operation. Longer durations indicate a user is more focused on a task or more emotionally stable; shorter durations may reflect a quick decision-making process, potentially associated with inattention, tension, or low anxiety. Individual differences in this characteristic may contribute to differences in attention span, finger control precision, and emotional state.

[0026] Speed ​​Change Mathematical symbol representation Where vi is the velocity of each point. Biometric Information: Speed ​​variations reflect the instability of the user's swiping speed. Large speed variations indicate significant emotional fluctuations, a more urgent decision-making process, or significant tension. Individual differences in nervous system response, motor control ability, and emotions such as anxiety may contribute to varying speed variations. Smaller speed variations indicate smoother, more relaxed, or more skilled operation.

[0027] Dispersion Mathematical symbol representation Among them, vi is the speed of each point, and avg_speed is the average speed. Biometric Information: The degree of dispersion reflects the fluctuation in the user's sliding speed. Higher dispersion indicates greater speed variation during the operation, which is often associated with unstable operation, anxiety, or lack of precise control. Lower dispersion indicates more stable and precise operation, which is often related to the individual's muscle coordination, neural control ability, and psychological state (such as calmness or concentration).

[0028] See also Figure 2 , Figure 2It is a VAE (Variational Autoencoder) inverse data trainer based on the owner's data. The algorithm reduces the dimensionality of the input owner user data by constructing an Autoencoder neural network. After standardizing the 20-dimensional owner user secondary data, it is mapped into a 10-dimensional latent space. The space cannot be too high or too low, otherwise it will affect the efficiency and accuracy of the model dimensionality reduction. It is then raised back to its original dimension through the decoder neural network. The neuron weights of the hidden layers of the encoder and decoder neural networks are adjusted through the back propagation algorithm according to the difference between the decoded data and the standardized data through the reconstruction loss and KL divergence function. At the same time, the val loss value is output to reflect the effect of a single training. The process is repeated many times to achieve the purpose of training the model. When the training result stabilizes at the minimum value, the model weight file VAE_MODEL.h5 and the standardized model Scaler.pkl are output and saved.

[0029] See also Figure 3 , Figure 3 This is a VAE (Variational Autoencoder) inverse data generator based on the owner's data. This algorithm constructs a neural network encoder and decoder with the same structure as the trainer model, reads the model weight file and normalized model data to create the model trained by the trainer, then inputs the owner's secondary data into the encoder for dimensionality reduction to a 10-dimensional low-dimensional latent space. The probability density distribution of this portion of the low-dimensional space is then calculated, and random points are taken from the low-density portion of the probability density distribution within the high-density neighborhood. The obtained points are then upscaled using the decoder and subjected to an inverse normalization operation to restore the scale to the same as the original secondary data of the owner. Finally, a CSV file with the same format as the owner is generated. This achieves the purpose of generating data for non-owner users. (Attached is a two-dimensional distribution diagram for auxiliary explanation.)

[0030] By training the VAE encoder and decoder, the corresponding decoded high-dimensional data in the latent space after data dimensionality reduction is extremely approximately equal, thereby achieving the purpose of approximate dimensionality increase. The approximate dimensionality increase technology is used to increase the dimensionality of the data points sampled in the low-density area of ​​the latent space to obtain an inverse data set obtained by increasing the dimensionality of the data in the non-user model other than the low-dimensional model of the owner user data.

[0031] Encoder structure: three hidden layers, 512 neurons in the first layer, 256 neurons in the second layer, and 128 neurons in the third layer.

[0032] Decoder structure There are three hidden layers, the first layer has 128 neurons, the second layer has 256 neurons, and the third layer has 512 neurons, which is the opposite of the encoder structure.

[0033] VAE sampling and loss function adjustment neuron weight principle (back propagation adjustment principle and training principle) forward propagation: the input data x passes through the encoder to obtain the mean μ and standard deviation σ of the latent variable. Through the reparameterization technique (Reparameterization Trick), sample z = μ + σ·∈, where ∈ ~ N(0, I) (Since VAE needs to sample the latent variable z, but the sampling operation is not differentiable, directly calculating the gradient will make training impossible. Therefore, VAE uses the reparameterization technique to split the random sampling into a deterministic part (mean and standard deviation) and a random part (standard normal variable ∈∈), thereby allowing the gradient to be back-propagated through μ and σ, avoiding the non-differentiable problem caused by sampling.) Send z to the decoder to generate the reconstructed data x′

[0034] Loss Calculation Reconstruction Loss: This part measures the difference between the reconstruction output of the decoder and the original input data. It is usually calculated using mean square error (MSE) or binary cross-entropy. Here, q(z|x)q(z|x) is the output distribution of the encoder, which represents the probability distribution of the latent variable z given the input x; p(x|z)p(x|z) is the output probability distribution of the decoder, which represents the probability of reconstructing the input data given the latent variable z.

[0035] KL Divergence: This part measures the latent variable distribution of the encoder output The KL divergence is the difference between q(z|x)q(z|x) and a prior distribution (usually the standard normal distribution N(0,I)N(0,I)). The goal of the KL divergence is to make the distribution learned by the encoder as close as possible to the prior distribution, thereby ensuring that the latent space has good structure and interpretability. Among them, μ j and is the mean and variance of the latent variable distribution output by the encoder, d z is the dimension of the latent space.

[0036] Back propagation: using gradient descent calculation formula Gradients with respect to encoder parameters The encoder of VAE outputs the distribution parameters of the latent variable zz, namely the mean μμ and standard deviation σσ (or logarithmic variance logσ2), so the gradient needs to be derived with respect to these two parameters.

[0037] Gradient of KL divergence with respect to mean and standard deviation:

[0038] Reconstruct the gradient of the loss with respect to the mean and standard deviation (calculated by backpropagation): Calculation by the chain rule

[0039] Gradient with respect to decoder parameters: in: Calculated by a loss function (such as mean squared error or binary cross entropy); Computed by the decoder's neural network weights.

[0040] Gradient descent update formula Encoder parameter update Decoder parameter update in: η is the learning rate. θE is the parameter of the encoder. θD is the parameter of the decoder.

[0041] See also Figure 4 , Figure 4This is a trainer based on the FCNN fully connected neural network architecture. This algorithm was developed using the TensorFlow deep learning framework developed by Google. By building a trainer based on a fully connected neural network (FCNN), it trains classification on the input secondary data of the user and the owner. The 20-dimensional secondary data is first normalized, and all features are mapped to a distribution with mean 0 and variance 1 to ensure rapid convergence and avoid numerical instability during training. The normalized data is then fed into a three-layer FCNN: the first layer contains 5 neurons and uses ReLU as the activation function; the second layer contains 3 neurons and also uses ReLU activation to better extract deep features; the final output layer contains 2 neurons and uses a softmax activation function to output a binary probability distribution. To prevent overfitting, a dropout layer is added to the network and an earlystopping callback function is introduced to monitor the val_loss of the validation set. Training is terminated when the val_loss stops decreasing to prevent the network from performing too well on the training set but failing to generalize to the test set. At the same time, cross-validation (KFold) is used to divide the training data into multiple folds and perform multiple training evaluations. Each fold generates an accuracy, and the average of the multiple fold results is used to measure the overall stability of the model. After the training process converges, if this is the first training, the trainer network weight file is saved as model.h5, and the scaler object used for normalization is saved as scaler.pkl. This allows new data to be normalized using the same parameters and maintain a consistent feature distribution. If a model file already exists, the previous network weights and scaler are loaded, and the trainer model is trained for an additional 50 epochs to further optimize classification performance.

[0042] See also Figure 5 , Figure 5This is a trainer based on the FCNN fully connected neural network architecture. The discriminator model has a similar structure to the trainer, differing in that its primary function is to identify and discriminate between user and non-user input data. In practice, the discriminator first reads the saved model weight file (model.h5) and the normalized model file (scaler.pkl), allowing for direct inference without retraining or minimal fine-tuning. Any new input data is first normalized using the same scaler.pkl file before being fed into the discriminator model for forward computation, resulting in a probability distribution indicating whether the sample belongs to the owner or non-owner category. The final discrimination result is obtained through operations such as argmax, and is recorded or processed based on actual needs. In this complete workflow, the trainer and discriminator, respectively, are responsible for model construction and training and final discrimination. Working together, they can quickly identify samples with significantly different characteristics from the owner, achieving efficient and stable binary classification.

[0043] Both the trainer and the discriminator use the same three-layer FCNN structure and its Dropout anti-overfitting strategy, supplemented by callback functions such as EarlyStopping and ModelCheckpoint to dynamically monitor training quality and automatically save the optimal model weights. At the same time, standardization, cross-validation, and multiple training evaluations are used to ensure consistent performance of the model on the training and test sets, making the entire binary classification system more robust and scalable when processing samples of both owner and non-owner users.

[0044] Forward Propagation: Overall process: For a fully connected neural network (FCNN), assuming the input sample (Here n = 20), the network consists of several layers, and the calculation process of each layer is as follows:

[0045] Linear transformation: For the lth layer, its input is a (l-1) (Layer 0 is the original input x), the weight matrix is ​​W (l) and the bias vector b (l) , then: z (l) =W (l) a (l-1) +b (l)

[0046] Activation function: After nonlinear transformation, the output is: a (l) =f(z (l) ) In the code, the first two layers use the ReLU activation function: ReLU(z)=max(0,z) The last layer uses the softmax activation function for binary classification: Output vector a (L) represents the predicted probability of each category.

[0047] Loss function: Categorical Cross-Entropy For one-hot encoded label y and predicted probability The cross entropy loss function is defined as:

[0048] Cross validation (K-fold cross validation): Divide the training data into K folds, use K-1 folds for training each time, and use the remaining 1 fold for validation. Let the accuracy of each fold validation be Accuracy k , then the average accuracy of the overall cross-validation is:

[0049] The backpropagation algorithm is used to calculate the gradient of the loss function with respect to the parameters of each layer, thereby updating the network weights. The specific process is as follows:

[0050] For the output layer, the error (gradient) can be written as: δ (L) =a (L) -y

[0051] For the lth layer (l=L-1,L-2,…,1), according to the chain rule: δ (l) =(W (l+1) ) T δ (l+1) ⊙f′(z (l) ) Where ⊙ represents element-wise multiplication, f′ is the derivative of the activation function, for ReLU:

[0052] The gradient calculation formula of the parameter is:

[0053] Weight updates are usually performed using the gradient descent algorithm (or its variants, such as Adam): Where η is the learning rate.

[0054] Prediction Category:

[0055] Accuracy definition:

[0056] Measures to prevent overfitting: Dropout is a random deactivation strategy that sets some neurons in the hidden layer to 0 with probability p (usually p≈0.2 to 0.5) during training, thereby forcing the network to not rely on local features: where r (l) ~Bernoulli(p) is a random vector that follows a Bernoulli distribution. At test time, no dropout is performed, but the desired consistency can be maintained by scaling the activations (or dividing by p during training).

[0057] Judgment algorithm: After forward propagation of the network, the output layer gives a probability vector: Judgment process:

[0058] The above description is only an implementation method of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this application, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An authentication method based on touch data of an Android smart device, characterized in that: The authentication method is applied to a data security system and includes the following steps:

1. Data collection steps: Use a specific algorithm to obtain real-time data on a single touch action of an Android smart device user. Using a single action as the sampling unit, collect the horizontal pixel coordinate (x), vertical pixel coordinate (y), touch area (s), and time from the start of the action (t) every 0.008 seconds. The collected raw data is then exported to a CSV file and stored in the specified address A.

2. Data naming rules: Use specific naming conventions for the output CSV files, and the file name format is Touch_data'x'('y'), where x represents the touch action number and y is the generated file number. 300 files are generated every 2.4 seconds, and the condition x≤y≤300 is met. When y reaches 300, both x and y are reset to 1 and the counting starts again.

3. Data management steps: adopt an approximate sliding window deletion strategy for the data files in address A, with 100 files as the interval unit. When the value of y reaches n×100 (n∈(1,2)), delete the files from [(n-2)×100+1 to (n- When y reaches n×100 (n=3), delete the files from [300+(n-2)×100+1 to 300], keep the storage space occupied no more than 500kb, and extend the single file reading time to 0.8 seconds; 4. Feature Engineering: Read the raw data from address A and calculate it using a specific mathematical formula to form a high-dimensional feature dataset. Generate a high-dimensional feature data CSV file for each action and store it in feature address B. The file name format is calculated_data('a'), where a represents the action number.

5. Data cleaning: Regularly traverse feature address B and delete all invalid files with a distance value of 0 in the feature data to ensure the quality of the dataset; 6. Non-owner data generation: A deep learning model is built based on high-dimensional feature data. The encoder maps the owner user data to a low-dimensional space. Directed sampling is performed in the low-dimensional space to generate non-owner user data. The decoder maps the data back to the high-dimensional feature space and outputs it to feature address C. The file format and naming are consistent with the owner user data.

7. Data merging: Merge the owner user data in feature address B with the non-owner user data in feature address C and store them in feature address D to form a data set for model training.

8. Neural Network Training: A fully connected neural network model is constructed. The high-dimensional feature data of the device owner and non-owners in feature address D and their corresponding labels are input for supervised learning training. Model parameters are optimized through backpropagation, ultimately forming an authentication model based on low-dimensional biometrics to effectively identify and distinguish between device owners and non-owners.

Images