An Adaptive Context-Aware Authentication System on Smartphones Using Machine Learning

Recently, the smartphone has been equipped with various sensors to enable users to unlock the smartphone screen. Authentication to unlock the screen is the mode that occurs when the user turns on the smartphone or wakes it from sleep mode. To remove the lock, the user must use one of the supported authentication methods. Authentication refers to verifying a user’s credentials, and authorization occurs after the system successfully authenticates the user’s identity [5, 16]. 

2.1 Symbolic password

This method of unlocking an Android-powered device requires the user to enter a combination of letters, numbers, or characters. The longer the password, the higher the security level of the smartphone. Users typically prefer passwords that are highly guessable, write them down, never changed, or only be changed among a few alternatives, making them vulnerable to attack. The password should be unique and memorable, and at the same time, it is easy for an unknown person to spy on the password the moment the user unlocks the screen [1, 2]. In this method, the user can always type the letters, numbers, and characters to create a complex password to unlock the screen. While this can be more secure than a PIN code or a pattern. Also, it would be very annoying to do it over and over again, especially, when the user’s hands are busy with something.

2.2 PIN code

As with the symbolic password, the PIN code enables the user to quickly unlock the smartphone. To unlock an Android smartphone or tablet, the user needs to enter a four-digit code. A PIN code is a reasonable authentication method that balances convenience and security. Users can type a longer PIN code faster than with a drawing pattern, and much faster if a simpler keypad is used. PINs might be guessed, shared, or robbed, especially if they are not kept confidential. For the anonymous, it is easy to spy to get the code. One of the main drawbacks of PIN codes is the limited level of security, particularly when compared to more recently developed authentication methods such as biometrics (fingerprint or face recognition). For a PIN code, users tend to pick the code that is easy to guess. Most PIN codes are relatively short, usually four to six digits long, making them vulnerable to brute force attack [1, 5].

2.3 Graphic or pattern lock

Pattern lock has become a popular method to authenticate a smartphone’s user. In this case, the user needs to draw a simple shape with his/her finger. It is a convenient and secure method to unlock the smartphone or tablet. The pattern is usually a 3X3 matrix of dots. The pattern method seems to be more secure than the conventional PIN code and less secure compared to biometric methods. However, swiping a complex pattern takes longer than a PIN code, therefore, the more secure your smartphone is, the more difficult it is to unlock as well [1, 2]. Patterns are easy to guess or notice, especially if the attacker watches the user while drawing the pattern.

2.4 Fingerprint

Fingerprints are intricate, almost unique, difficult to change, and persist throughout an individual’s life, making them appropriate as a long-term feature of an individual’s identity. Fingerprint verification is a process for identity authentication used to verify an individual’s identity by comparing a captured sample with a previously stored sample. When the samples match, the verification process is passed. In this case, the user’s fingerprint pattern is only recorded once, which makes it easy to verify the fake fingerprint successfully. Fingerprint can be used to open the lock easily and quickly. It is a reliable security method that is hard to tamper with. The user needs to touch the sensor with his/her finger. It also does not require the genuine user to remember a password. Unlocking has become much easier with the advent of fingerprint sensors in smartphones [2, 17].

2.5 Face recognition

This method enables genuine users to unlock their smartphones and verify their identity simply by placing their faces directly in front of their smartphone screen. Face images are used to build the user’s face model, and the face model is stored on the smartphone. Facial recognition is a secure and convenient method to unlock a smartphone device. This method uses facial recognition technology to identify the owner of the smartphone. It takes and analyzes the owner’s unique facial features to build a template of the face. During the setup process, the smartphone owner will be asked to record his/her face by taking a sequence of images. Then, the smartphone builds a model of the owner’s face that can be used in the comparison process. Although the facial recognition method is convenient, it is not secure as compared to other methods such as fingerprint or a PIN code. For instance, some systems can be tricked by models or videos of the user’s facial features [6, 18].

2.6 Voice unlock

This method allows genuine users to save voice samples on the smartphone and when the user speaks the words loudly, the smartphone will check the sound to unlock the screen [18]. Voice unlocks systems recognize and verify the smartphone owner’s voice by analyzing various voiced traits, such as tone, pitch, and speech prints. Although voice unlock is convenient, it is not secure as compared to other biometric methods such as fingerprint or face ID. An attacker can impersonate the owner or record the owner’s voice. Voice unlocks is often used for hands-free access to smartphones. Users can just say a passphrase to quickly unlock their device. It is a very appropriate approach, especially when the user’s hands are busy eating or driving. Regarding privacy concerns, the genuine user should be aware of potential concerns as voice samples are usually kept and processed on the device [18].

Certainly, smartphone unlocking methods differ in terms of security, convenience, and accessibility. A comparison of common smartphone screen unlocking methods highlighting the strengths and weaknesses of each method is illustrated in Table 1.

Table 1. A comparative summary of smartphone unlocking methods

In this section, a framework architecture for an adaptive context-aware mobile authentication system is presented. The proposed framework consists of a set of layers that can sense context changes and adapt the system’s authentication model according to the context as well as the user profile and preferences. 

Figure 1 provides an overview of the context-aware authentication system, which consists of two stages: training and authentication. In the training stage, the proposed system calls the motions, location, and other sensors built into the smartphone to gather the owner’s data when he wants to unlock the smartphone screen. In addition, from the collected data we will extract contextual features and define user context according to the contextual awareness classifier. The detected context features are fed into the machine learning model to provide an authentication method. 

1.jpg

Figure 1. Architecture of the proposed authentication system framework

In the authentication stage, the user just needs to unlock the smartphone screen as usual to finish the authentication process. During system execution, the obtained features and captured context information are provided in the trained proposed model. In the next part, we will describe the main components of the proposed approach, namely data acquisition, context awareness, machine learning model, and authentication model construction.

4.1 Context sensing layer

The first step is to obtain the user’s input data. After the transformation of the context, the second step generates the customized outcome according to the information that reflects the authentication model. Contextual information can be categorized into seven elemental kinds: geographical location, alarm settings, calendar events, SMS messages, noise levels, weather conditions, and user’s physical activity as shown in Figure 2. 

2.jpg

Figure 2. Types of contextual information

In our mobile authentication system, contextual information originates from manifold physical and virtual sensors. The component performs a user-centered operation using smartphones. Smartphones can use context information to render a personalized authentication model. Virtual and physical sensors can be used to obtain the user’s context information as follows [21]: 

4.1.1 Virtual sensors

Virtual sensors can obtain virtual information from mobile apps. For instance, applications such as Calendars, Alarms, Weather, and SMS can be deemed as virtual sensors. Specifically, we can use the Alarm App to get a wake-up alarm and sleeping schedule for the user. In our model, the user’s virtual sensor contexts are composed of four factors: calendar events, weather conditions, alarm settings, and SMS messages. Reading from virtual sensors can involve different potential challenges, including:

• Calendars: Scheduling conflicts, time zone differences, event updates, privacy concerns.
• Weather conditions: Inaccuracy, changing conditions, regional variations.
• Alarm settings: Alarm types, privacy and security, setting customization.
• SMS messages: User permissions, message formats, message encoding schemes, length limitations, privacy concerns.

4.1.2 Physical sensors

The Android smartphone is equipped with sensors that detect movement, direction, and various environmental states. Android sensors relay data from various physical sensors such as accelerometers, ambient temperature, magnetometers, gyroscopes, heart rate, light, proximity, pressure, and relative humidity sensors. Reading data from physical sensors can come with several potential challenges. Here are some common problems associated with these sensors:

• Location sensor: Inaccuracy, signal interference, indoor use, battery drain, privacy concerns, real-time updates, etc.
• Accelerometers sensor: Noise and signal filtering, vibration and shock handling, power consumption, calibration, etc.
• Gyroscopes sensor: Noise and drift, temperature sensitivity, vibration and shock, power consumption, latency, calibration, etc.
• Noise level sensor: Accuracy and calibration, environmental interference, privacy concerns, sensor location, stable power supply, etc.

The proposed approach utilizes three broad categories of sensors as follows [22]:

4.1.3 Motion sensors

These sensors (including accelerometers, gyroscopes, gravity sensors, and rotation vector sensors) are used to monitor the user’s physical activity such as immobile, walking, running, etc. The accelerometer is used to measure the appropriate acceleration or precise movement of the smartphone and detect the shake and inclination of the device. A gyroscope is a sensor that can measure and maintain the rotational rate and angular velocity of a smartphone. The rotation vector sensor can be used to detect gestures and measure relative changes in the smartphone’s orientation. Thus, after the user’s context information is collected by the smartphone, the proposed approach will segment the context information according to the source sensor type.

4.1.4 Environmental sensors

The smartphone provides sensors to measure different environmental attributes. These sensors are used to monitor the ambient temperature, relative humidity, illuminance, noise level, and pressure, in which the user is operating and located. All environmental sensors are hardware-dependent and are only available if the smartphone manufacturer has included them in a device.

4.1.5 Position sensors

These sensors are used for monitoring the physical position of a smartphone device (e.g., home, office, library, etc.). Android-powered devices use two sensors to determine the location of a smart device: an accelerometer and a geomagnetic field sensor.

4.2 Context feature extraction layer

In this layer, context feature extraction is responsible for tuning and analyzing contextual information collected by virtual and physical sensors. The raw sensor data gathered by these sensors are processed in this layer, and the output information is retained for use in the machine learning layer. The goal is to determine the appropriate context for our model requirements and the degree of context relevance that matches the user’s context.

To perform better user authentication, we use the extraction method to extract features from the obtained user data. In this system, context features are heterogeneous and from various sources (i.e., virtual and physical sensors). We need to define the dataset in machine learning, how to gather the data, and what features the appropriate dataset has. That is why, after data acquisition, it is essential to pre-process it (data cleaning), as well as annotate the data by labeling it so that can be understood by the system. In addition, we suppose that the user’s context when using the smartphone will be helpful for authentication model construction.

Similarly, in our approach, motion sensors (accelerometer, gyroscope, and rotation vector sensor) are used to get the motion attributes of the smartphone while unlocking the screen. Moreover, by analyzing the user’s context information while unlocking the screen, it can be seen that the uniqueness of various users’ context is not only deemed in the motion sensing information of the smartphone but also in the location information, ambient environment state (noise level), and software context (i.e., weather condition, calendar events, and alarm settings).

4.3 Machine learning layer

Machine learning is a key component of the proposed approach. Using statistical techniques, algorithms are trained to perform decision processes (i.e., prediction or classification). A machine learning model is a mathematical function that takes features as inputs, makes predictions as outputs (the predicted value is the label), and learns how best to correspond predicted values with patterns reported from the training data. In machine learning, the learning process is just a mapping function that maps features to targets and this function is used for data whose labels are not known. In the proposed approach, the authentication model is one of the most critical parts for selecting the adaptive and most related authentication method [13, 23].

4.3.1 Multiple linear regression

Multiple linear regression is a machine learning technique used to model the relationship between two or more independent features (i.e., inputs or predictors) and responses to one dependent variable (i.e., the output). MLR is a statistical model that produces coefficients that can be used to predict inputs by fitting a linear equation to the observed data related to each user. MLR is used when the system needs to predict the outcome of a variable based on the value of other independent variables. It is an extension of linear regression models that takes only one independent variable. In MLR, the target is to understand how the dependent variable is affected by two or more independent variables. It can be defined using the following formula [14]:

$y=\mathrm{R}_0+\mathrm{R}_1 X_1+\cdots+\mathrm{R}_n X_n+\partial$     (1)

where:

y =The dependent, i.e., predicted response.

$\mathrm{R}_0$=The y-intercept, i.e., the value of y when all other factors are zero.

$\mathrm{R}_1 X_1$=The coefficient of regression ($\mathrm{R}_1$) of the independent input variable ($X_1$) (It is also known as the impact of increasing the value of the independent variable on the predicted value of y).

$\mathrm{R}_n X_n$=The coefficient of regression for the last independent input.

$\partial$=A model error (that is, how much variance is there in the estimation of y).

Regression is used to find relationships between independent variables as well as map features to others. The proposed model attempts to find out how users’ authentication methods depend on their features. We use multiple regression to predict the authentication model based on the user’s physical activity, geographical location, weather conditions, and calendar information. With multiple regression, we can use more variables, such as alarm settings, and SMS messages to make the prediction more accurate [14].

4.3.2 Dataset

The first thing we need to do is decide which sources we will use to collect data. Typically, there are three types of sources you can select from: open-source datasets available online, the Internet, and artificial data generators. Dataset collection sources vary and depend heavily on the targeted project. The best option for our model is to collect data from smartphone sensors that are directly related to the system objectives [24]. 

Datasets are especially essential for machine learning algorithms to learn from. The dataset is an instance of how a machine learning algorithm can make predictions, with tags representing the result of a given prediction (pass or fail). A dataset is just a collection of data that can be utilized to make predictions about future events or outcomes based on historical data. Datasets are usually labeled before being used by machine learning algorithms, so the algorithm understands which results to predict or classify [14, 24]. 

Raw data is a useful source for starting and feeding a machine learning algorithm to deliver users with an appropriate authentication model according to the user’s context. Machine learning datasets take various forms and can be obtained from several sources. Typically, the three kinds of machine learning datasets are text data, multimedia data, and sensor data.

Furthermore, it is necessary to make sure that the data fragments are of sufficient quality. While there are manners to clean up the data and make it standardized and understandable before the two processes: annotation and training, it is best that the data matches the list of desired features. There are a few steps we need to perform before the dataset can be usable [25] (see Figure 3).

• Collect: The first thing we need to do is identify data sources. In our case, the raw sensor data is the source of the dataset for the proposed approach.
• Pre-process: We need to define which features make up a suitable dataset. High quality is the key thing to consider when collecting a dataset for a machine learning model. First, the pieces of data must be relevant to our objective. Second, the data should be cleaned by fixing bad data in the dataset, including, empty cells, wrong formatting, duplicates, etc.
• Annotate: After we make sure that the data is clean and relevant, we also need to ensure that the data is understandable so that the computer can process it.

Thus, to make smart decisions and deliver an adaptive, secure, and context-aware authentication model, context awareness is now a crucial requirement for mobile computing and services. To build our model using machine learning, we must create a new dataset. Because it is a very specific problem, we must create a domain dataset, clean it, and understand its relevance to acquire the outcome. Next, it is time to train the proposed model. The basic principle is to get the annotated data and feed it into the algorithm. In fact, we will need to repeat this process several times. This step is a training of the machine learning model. Training means feeding our proposed model with bunches of cleaned and annotated data for a distinct number of epochs (i.e., the number of times the algorithm is trained on the entire project).

3.jpg

Figure 3. Three steps to processing data in machine learning

4.3.3 Machine learning features

A suitable dataset is required for the machine learning model; therefore, we must determine the features of the collected data. For this reason, physical and virtual sensors are chosen to collect the following contextual factors.

4.3.4 User’s body postures and physical activity

Smartphones are not always immobile, and users may use their smartphones while they are in various body postures and physical activities, such as running, walking, sitting, or driving. The context of physical activity refers to the user’s actual postures. For example, an increase in smartphone mobility is less favorable toward a PIN code or password authentication model, and thus, face recognition is more favorable. According to the user’s mobility, we classify the user’s posture and activity into four levels: sitting, walking, running, and driving. Each level is defined by a numerical value as shown in Table 2.

Table 2. User’s posture and activity levels

4.3.5 Noise levels

The smartphone is used to compute noise levels in the surrounding environment such as crowded spots in cities or workplaces, and quiet places in the countryside or at home. When the ambient environment noise is loud, voice-based authentication is not usable. Noise levels affect the user’s concentration regarding achieving authentication. The factor will be used as one of the machine learning features. Three levels of noise are defined as shown in Table 3.

Table 3. The values of noise levels

4.3.6 Geographical location

In the proposed authentication model, the geolocation context, defined by the location sensor, is an influential feature in providing acceptable authentication as we assume. This approach uses location-tracking technologies to determine the owner’s physical location. For example, in a library, the fingerprint model is more favorable to be delivered, or a user can unlock her smartphone once she enters a certain location. We classified an individual’s physical location into four categories. They are expressed in integer values as shown in Table 4.

Table 4. Physical location parameter values

4.3.7 Alarm App settings

Users usually set a wake-up alarm and sleep schedule (including bedtimes, wake-up times, and mealtimes) in the Alarm App. If the user does not want to set up a sleep schedule, he/she can set a regular alarm for the wake-up time. For instance, when a user is having a meal, a voice-based authentication method is more convenient. We categorize an individual’s sleep schedule into three categories as shown in Table 5.

Table 5. Alarm settings values

4.3.8 Calendar App events

The Calendar App is a storehouse of user events. The Calendar App automatically synchronizes any events and dates linked with the user’s Google Account. On Android devices, the Calendar application displays events and is most typically used to set reminders for events. However, when the calendar event is a meeting or conference, the fingerprint method is preferable. Two types of Calendar events and their corresponding values are identified in Table 6.

Table 6. The values of calendar events

4.3.9 Weather conditions

Nowadays, weather forecast apps are an important part of smartphones and are one of the most frequently visited apps. The proposed approach uses a Weather App to check the weather conditions. We can get various details like temperature, precipitation, humidity, cloudiness, and so on. For example, the fingerprint sensor cannot read your fingerprint if your hand is wet, damp, or oily, so facial or voice-based recognition is more convenient. However, in our model, we considered three different types of weather including Sunny, Rainy, and Snowy as shown in Table 7.

Table 7. Types of weather values

4.3.10 SMS App

Today, any smartphone has SMS capabilities and can send and receive text messages to or from any other smartphone through the messaging app. In the proposed model, when the smartphone is stolen or lost, the genuine user sends the following SMS message to his/her smartphone “My phone is stolen or lost.” We considered two different situations of a smartphone including stolen/lost or existent as shown in Table 8.

Table 8. The two different situations of a smartphone

All the above features are important for the authentication model as they determine the authentication method to be selected.

4.4 Authentication model

An important component of this layer is the authentication model required to be defined. In the proposed model, we assume that the authentication model depends on the user’s current context. Using context factors, we designed an authentication model whose type is determined by the value of motion, location, and noise level sensors as well as the software context (Weather, SMS, Alarm, and Calendar Apps) as shown in Figure 4. At this point in our model, we assume that the type of any authentication model describes a combination of all context factors.

4.jpg

Figure 4. The process of designing an authentication model using machine learning algorithm

To find out features from the raw data, users are asked to go through the authentication process to unlock the smartphone screen so that we can get their responses according to the generated authentication model. To increase the accuracy of the proposed model, the dataset is updated each time the authentication process is performed. When a user inputs a PIN code, draws a pattern, and enters a fingerprint, the proposed model records the user’s response using emoji sentiment (see Figure 4). Six methods of authentication methods and their corresponding values are identified in Table 9.

Table 9. The value describes the authentication method

Machine learning algorithms are only as useful as the data used for training. The more data samples used during training, the more useful your algorithm will be. This is why it is necessary to have a large volume of processed datasets when working on artificial intelligence systems– so you can train the model effectively and gain the best outcomes. When we deal with machine learning models, data is fundamental. Without data, we cannot perform typical model training, and no insights are achieved.

We need to build a model that can predict the most appropriate and secure authentication method for a smartphone based on the user’s context. In our case, the target variable (or label) is the authentication method, and the features are the user’s body postures, location, noise level, weather conditions, calendar events, SMS messages, and alarm settings.

We will collect data from different users when they unlock their smartphone screens, with their corresponding inputs and outputs. This data is labeled and it’s in the form of a user’s context for each user. Pre-processing is performed to fill in any missing values or address outliers that might impact the accuracy of our model.

Depending on the user’s context and surrounding information, the appropriate authentication model is selected, adapted, and delivered to perform the authentication process. For instance, the increase in smartphone mobility is less favorable towards a PIN code or password authentication model, so, when the user is running or driving, the most appropriate authentication method will be facial recognition or voice unlock respectively. 

The smartphone detects the user’s location and changes its authentication method, for example, automatically changing to pattern mode if the user is in a meeting, changing to fingerprint if the user is at home, and changing to voice ID whether the user is walking, running, or traveling by car. The smartphone detects whether a place is crowded such as an airport, railway station, or mall, and automatically changes the device authentication method to provide a non-voice command model.

However, if the smartphone is stolen or lost, the attacker may utilize lighting conditions to unlock the smartphone by picking up smudges remaining on the screen by the genuine user. Therefore, in our model, we add an extra level of security along with the password-based authentication method. We have implemented three levels of authentication. The first level is the symbolic password, the second one is the user’s fingerprint, and the last stage is facial recognition. Therefore, by going through these stages, the genuine user can be identified, and access granted. 

6.1 Participants

To collect data, 25 participants were invited to participate in the study. Participation is entirely voluntary. All participants are smartphone savvy and use authentication mechanisms to protect their personal information. Participants ranged in age from 18 to 45 years, with a good background on the direction of the study. The demographics of participants are shown in Table 10. These participants were asked to operate on their Android smartphones. Participants in a research study were given instructions via email, and these instructions were designed to ensure that participants fully understood what was expected of them. In addition, we take any ethical issues into account to protect user privacy.

Table 10. Demographic characteristics of the participants

6.2 Authentication methods setup

To evaluate the effectiveness of the proposed system, every time users turn on their smartphone or wake up the screen, they will be asked to unlock their smartphone, therefore, each participant needs to set up the screen lock in the following ways: 

M 1: Pattern; Draw a simple figure with user’s finger.

M 2: PIN; Enter four or more numeric code.

M 3: Password: Enter four or more characters or numbers.

M 4: Voice; Detects the user’s voice and unlocks the smartphone when the user says the correct password.

M 5: Fingerprint; The sensor reads the user’s fingerprint pattern.

M 6: Face ID; Takes images of user face from different angles. The face model is stored on smartphone.

6.3 Operational environment

Because this study is based on context-aware authentication, we consider different operational environments to collect data. To verify the relevance of the authentication model based on contextual awareness, we examine the performance of authentication in various usage contexts. Authentication experiments are performed in the following physical activities:

A1: Training sets are obtained from the data collected while a user is in a non-moving position.

A2: Training sets are obtained from the data collected during the user’s walk.

A3: Training sets are obtained from data collected while the user is running.

A4: Training sets are obtained from data collected while the user is driving.

To make the collected data realistically reflect the user’s attitude habits, all participants are required to use the smartphone at different time intervals: alarm information (bedtimes, wake-up times, etc.) and calendar events (meetings, conferences, etc.) with various weather conditions (rainy, sunny, etc.). Furthermore, the 25 participants need to use a smartphone to authenticate themselves while they are in different physical locations (home, office, park, etc.) and at different noise levels.

Collecting contextual data (i.e., personal or private information) may present ethical issues and needs compliance with privacy rules. The proposed system ensures that users know what data is collected, how it is used, and that they have provided their consent to data collection. However, the data is collected only for the specific goal for which individuals are informed. The system avoids using contextual data for unapproved purposes.

6.4 Evaluation criteria

6.4.1 Sentiment analysis

Due to the widespread popularity of emojis, many researchers have developed ways to integrate them into existing sentiment analysis methodologies [15]. The proposed model allows users to express their views and opinions using the most popular form of expression (i.e., Emojis). Each opinion expressed has an associated emoji called a sentiment (see Figure 7). Users were asked to provide their satisfaction or dissatisfaction regarding the authentication model. A total of 25 responses were collected as illustrated in Figure 8. 

7.jpg

Figure 7. The process of expressing opinions using emojis

8.jpg

Figure 8. A percentage value for each sentiment emoticon

We analyzed user’s satisfaction level with the proposed system. The sentiment ratio is calculated by comparing the number of negative sentiments emojis with positive or neutral emotions. From the analyzed data, we noticed that the sentiment ratio is 72 percent (see Figure 8), which means that context-aware technology is helpful in improving authentication adaptivity and provides a convenient authentication method.

6.4.2 Training and testing

To evaluate the performance of the proposed system, Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R Square metrics are adopted. More specifically, MAE is the absolute difference between actual and expected values. The lower the value, the higher the performance of the model, which is defined as follows [14, 24]:

$M A E=\frac{\sum_{i=1}^n\left|y_i-x_i\right|}{n}$         (2)

where, y_i are the expected values, x_i are the actual values, and n is a total number of data points.

Whereas RMSE is the standard deviation of the errors that occur when making a prediction on a dataset. The RMSE is taken into account when determining the model accuracy. The lower the value, the higher the performance of the model. The RMSE formula is defined as follows [14, 24]:

RMSE $=\sqrt{\sum_{i=1}^N\left(x_i-\widehat{x}_i\right)^2}$        (3)

where, N is the number of non-missing data points, xi is the actual time series of observations, and $\widehat{x}_l$ is the estimated time series.

Finally, R-squared (R²) is a statistical metric that describes the proportion of variance in a predicted variable explained by an input variable. It is the determination coefficient. It informs us how many points lie on the regression line. R² is defined as follows [14, 24]:

$R^2=1-\frac{\text { Sum Squared Regression (SSR) }}{\text { Sum of Squares Total (SST) }}$      (4)

where, SSR is the sum of the squares of the residual values, and SST is the sum of the distance the data is completely away from the squared mean.

The MAE obtained for the proposed model is 1.299, which means it is pretty good because it is close to zero. An MAE of zero indicates that the model is an excellent predictor of the outcomes. The calculated RMSE of the proposed model is 1.437, which is also pretty good. The R-squared value is 76.78, which means that 76.78% of the data fit the regression model as illustrated in Table 11.

Table 11. Performance of the proposed authentication system