A Secure Recommendation System for Providing Context-Aware Physical Activity Classification for Users

Department of Computer Science and Engineering, PSN College of Engineering and Technology, Tirunelveli, India School of Computing, SASTRA Deemed University, "anjavur, India School of Technology, Pandit Deendayal Energy University, Gujarat, India School of Computer Science and Engineering, University of New South Wales, Sydney, Australia Department of Management, Universitas Kristen Petra, Indonesia Department of Computer Science and Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology, Chennai, India


Introduction
Physical Activity (PA) is defined by the World Health Organization (WHO) as the body's movement using skeletal muscles while consuming energy. ese activities can be working, playing, performing domestic chores, wayfaring, and merrymaking activities. Routine PA demanding the average workforce like walking, cycling, or sports practices gives remarkable health benefits. However, inadequate PA is also one of the risk factors for worldwide mortality and keeps escalating in most countries, increasing the Non-Communicable Diseases' (NCDs) burden and affecting general health globally. e inadequately active people are at a high risk of death, 20%-30% compared to adequately active people. Many researchers and scholars from different spheres have been inclining towards health-oriented topics, commissioned by incorporating technological devices and tools in healthcare service platforms, called e-Health [1,2]. ese systems drive a growing number of users, and the stimulation is mainly due to the arrival and dispersal of smart devices and the pursuit of intelligent life. e outreach of the smartphone has been an excellent potential for a personalized Activity Recommender System (ARS) for individuals. Explicitly, many smartphone applications and wearable devices associated with Wireless Body Area Network (WBAN) track step count meticulously. Only a subtle difference was identified in the data collected from smartphones than the monitored step counts. However, the step counts could be either higher or lower. Simultaneously, the wearable devices showed huge discrepancies, and the step counts in 1 device were above 20%, which was relatively lesser than monitored.
Step counts are the usual means for measurement calculation of PA, such as distance or calories burned. e substantive difference in the precision of the device may be fused in these measures. ese devices eased increased PA, leading to clinical benefits unapprehend by low adoption of pedometers.
Usually, the users are endowed with elaborate interfaces by smartphone applications that visually track users' activity levels.
e MD and wearable devices used for the study proved to be very accurate for tracking step counts. e accelerometer data was used in other studies to recognize/ differentiate static and dynamic user states, day to day activities like running, leisurely walking, and riding bicycle [3,4], and some physical movements (sitting or standing, lying, walking, etc.) [5].
More Mobile apps to track physical activities have access to MD. e rapid development of cutting-edge Mobile Devices (MD) solutions such as Wi-Fi/GPS has prompted the development of Location-Based Social Network (LBSN), factors such as Loopt, Brightkite, Foursquare, Gowalla, and Whrrl, which subsequently increases the ability to evaluate to multiple locations as a novel trend for smartphone users [6,7], and it encourages users to leave comments, tips, or ratings of the visited sites/activities performed depending on their satisfaction level. All these facilities gradually expedite the growth of enormous data accumulation, which is the means to enhance user characteristics. Similarly, it facilitates experts in designing valuable services for users to compete with contemporary living standards.
Apart from the Point of Interest (POI) recommendation, which has received much attention in research and business [8,9], a great deal of services is being developed based on LBSN data that appear to be promising to users as well. At the same time, compared with various conventional encryption algorithms, there is a high level of confidentiality with a smaller key size. In the security analysis part, the suggested security strength [10,11] against various harmful security threats is demonstrated to ensure greater security. Recommendation Systems (RS) for outdoor activities such as playing, bowling, or strolling to an American fast-food place are examples of these services. ere is an increase in the performance of a variety of activities during leisure time.
e growing demand for e-healthcare systems requirements is due to increased technologies for a sporty and healthy lifestyle. Stability in an individual's health awareness is caused by an integrated health care system consisting of functional capabilities [12,12,13]. A headway in personalized technologies has elaborated the scope of research to achieve Digital Data-Driven Decision-Making Systems (DDDMS) connected with real-time health data. Digital health MD such as the Apple Watch, Google Fit, and Samsung Health, which accurately monitor personal Health and fitness tracking, is a key component of today's modern e-healthcare systems. ese devices are very supportive, as they suggest recommended features that efficiently analyze the human body, which include footsteps, heart rate, and hours of sleep [14][15][16]. ese systems additionally focus on fitness tracking.
ough Regular Physical Activity (RPA) benefits are testimonial, they have minimal effect in making inactive people cling to an activity. e surveys carried out in various industrialized nations show that 10% of inactive adults start a regular exercise program in a year, and 50% of the people who begin sports or RPA withdraw from the program in 6 months. Hence, increasing involvement or continuation of RPA has been a flourishing research area in sports medicine [17]. ough sufficient proofs exist regarding the reasons and the possible methods for increasing involvement, these proofs vary from country to country. For instance, self-efficacy is one of the leading causes of exercise adherence [10,18,19]. General Perceived Self-Efficacy Scale is a ubiquitous method for measuring self-efficacy. is has been used across diverse cultures and has been authorized through usual statistical analysis for diversified languages, including Japanese [20][21][22]. is paper proposes PRIPRO and a privacy-protected activity recommendation system that utilizes a Homomorphic Encryption (HE) model to protect user details and furnish accurate activity recommendations that depend on the user's present scenario. e paper is organized as follows: Section 2 consists of related works about various methodologies for detecting and classifying humans' healthcare. Section 3 describes the working principle and process of system structure and data set included in the proposed method, and Section 4 produces the best outcome of our model with the best accuracy. Section 5 shows the conclusion part, Section 6 acknowledges all the authors, and Section 7 gives the references we have referenced for producing the best result by overcoming the drawbacks of existing methods.

Related Works
Designing a DDDMS for exercise and sports involves many practical difficulties, complicating using one of the wellknown recommender systems [23,24]. Firstly, it is a different experience to see people doing the same exercise at the same places, not to mention differences between gymnasiums. An easy exercise, like walking, can even give new feelings when done at different spots. erefore, the proposed recommender system has been designed, so that people with a proven history of regular exercising and performing "similar" sports might also make "similar" inactive people active. In other words, the philosophy we have used in our design was the information associated with a lifestyle that we collected by doing complete health checkups in Japan and Ningendoku, which might be quite good indicators that explain how similar people are. If an inactive person wants to begin an activity, the best way is to discover a similar person and suggest the same exercise.
With the advancement of wearables in WBAN, like various Internet of ings (IoT) smart devices such as fitness trackers and smartwatches, the quantity and prototype of gathered health care data have expanded. As a result, this needs new and scalable structures to ensure user privacy and real-time medical data used for analytical purposes. is is extremely crucial, because the gathered multimedia data enable more number applications resulting in data leakage. For example, an average fitness tracker can collect information regarding the user's location, timestamp, heart rate, daily activities, nutrition, and sleep cycles. ese real-time health data are fetched from the whole user base and thoroughly analyzed for DDDMS to offer personalized recommendations. As the user real-time health data is wholly accessible by commercialized e-health solutions, this may become a hurdle for mass adoption of fitness devices, as DDDMS are driven by Machine Learning (ML) applications that are only beneficial by exploring huge volumes of medical data. Exploring the use of cryptographic techniques and differential privacy is done to enhance privacy in medical data communication systems and RS as a countermeasure [25][26][27][28][29]. Anyhow, to our knowledge, a bound privacy RS for the collection of real-time health data by MSs is mainly unexplored. e predominant challenge for designing privacy bound big-DDDMS, or health RS, includes the following (Table 1): (1) Opting for distributed and cost-efficient privacy maintenance solutions (2) e capability to back up a vast range of input data formats (3) Design of smart incentive models to motivate data sharing by users e PA-RS's biggest challenge is to make quality decisions by not compromising the active target user's privacy. Indeed, just about half of all individuals and a quarter of people over 65 years of age meet the minimum PA level needed to stay healthy (Dept. of Health in 2011). Inactivity is the leading cause of deterioration in physiological fitness and disease in the elderly. When the RS deals with a considerable quantity of sensitive user multimedia data, conventional recommendation approaches proved ineffective because of the generated data's scarcity and heterogeneous naturewith the evolving nature of the user's PA data, storing and computing the conventional data processing methods cost high. For decreasing computational overheads, it is highly suggested to shift the recommendation framework into a cloud environment. While getting into the cloud to produce quality recommendations, it is necessary to ensure user data privacy, especially when susceptible. erefore, there is a need for a secured RS to maintain the user's privacy with a fool-proof mechanism. For expressing the privacy and security problems of RS, a new encryption approach is warranted by not adding the computational costs. With the desperate need to protect heterogeneous user data, an entirely new HE is proposed considering the RS research requirements. e use of the cloud paradigm guarantees a trusted computational environment and produces secure and quality recommendations [30]. For IoT-based WBANs, an efficient and secure anonymous authentication architecture is designed with location privacy protection [31]. e detailed study demonstrates that the proposed approach overcomes the security flaws of existing schemes while simultaneously reducing computing costs. Figure 1 demonstrates the system structure of PRIPRO. Mobile Devices furnish users' data for application recommendations. A Mobile app is fixed in every mobile device, and various functional modules are part of this application. e function of Reliability Monitor is to monitor the user's usage patterns of Mobile app usage and input user's data regarding their Access Patterns (AP), Usage Time (UT), and Activity Data (AD) into Device Database (DD). e calculator performs necessary computations in MDs, e.g., summing AP, UT, AD, encryption, and decryption. Signature and anonymization of user identity are the responsibility of the Identity Manager. e Local Key Auditor performs the generation and management of related keys. Mobile data distributor passes encrypted data to cloudbased Recommendation Service Provider (RSP) and obtains data response. Device Database stores all the data in a secure way. RSP offers preprocessing data service (e.g., for recommendation purposes). RSP, basically a cloud service provider, possesses potential storage capacity, and computational capacity through user privacy is its primary concern. User Identity Manager (UIM) verifies the MD's ID and the processing of encrypted data correspondingly. e Privacy Protection Agency (PPA) manages smartphone access control and key management. Data Distributor enables the communication between MD and PPA in PPA. Figure 1 shows that the arrows present in each entity signify the flow of internal data. It is believed that MDs, RSP, and PPA communicate based on [32], which evaluates the application practice of users on their MD to find out their PA states. When the application detects a dormant state in the current environmental contexts, it recommends appropriate activities.

PRIPRO System Design.
e flowchart representing the RS comprises three parts: (1) Data collection, (2) Preprocessing, and (3) DDDDMS and recommendation, as shown in Figure 1. e description of each part is elaborated in the following sections:

Data Collection.
e Activity Recorder Module (ARM) facilitates the Data Collection Module inside the Mobile app to collect personal and sensor information, realtime medical data, and activity history. e first-time users of the proposed system are asked to fill out a questionnaire on the smartphone and maintain their personal information like age, gender, occupation, office hours, etc. e users would also fill out a scenario questionnaire that requires Security and Communication Networks them to select their interest activities under a different context. is paper defines context as a blending of environmental factors (e.g., office hours, workplace, and weather). Each Recommended Activity (RA) is recorded in the RS. e labelling of activity levels was 3D of the participant's PA: regular, occasional, and rare.
Further, each dimension was divided into three levels: Low, Middle, and High. When an activity is recommended, A novel framework for highly prevalent healthcare that manages to combine data transmission control and dutycycle adaptation.
Produces more energy consumption while processing through the on-WBAN channel.
T. Shaw et al. the user's current location (tracked using GPS/Wi-Fi) and time are detected to determine their environmental context. Activity item history maintains the user's current context and the selected activities (e.g., office hours, workplace, comfortable weather, music, etc.)

Data Preprocessing.
A real-time data must be preprocessed to spot the features inside a particular time window, which will be used as input for the DDDMS. e system's parameter binds the time windows, and it is unique for every user and every activity. e original data store records are categorized according to the user ID and activity in the first step. After that, the records are grouped as a timestamp in ascending order. e second step identifies the Jumps, which are durations when there is no data collection. e parameter of the framework is the jump time interval, which is to be found out experimentally. e jump time interval was fixed as 5 minutes in our framework. Lastly, the length of each interval was determined, which requires the identification of user activity. Study experimentation was done with specified time intervals of 15 Secs, 10 Secs, and 5 Secs.
More amount of data is needed to be gathered as it is generated every 50 ms. Apache Kafka is used to streaming the data in real-time, whereas Cassandra is ideal for storing time-series data, because it can handle real-time requests. e feature calculation mechanism ingests stored data in Cassandra, and later it enables the application's classifier algorithm to sort out the user's activity in near real-time. High scalability with more number of users is achieved in decoupling between data generation and classifier via Kafka. In the existing implementation, user activities' classification was carried out in one of the following categories: running, stepping up the staircase, swimming, skipping, and cycling. e classifier should classify each category with maximum accuracy, because some activities present the same characteristics (for example, running, and jogging). For identifying user activity for a specific time interval, we adopted an approach of getting inspired by the work of [33,34] and discovered the following features: e analysis of the features was done for specific window sizes, and later, the user's activity was classified according to the count for each window.

Recommendation Module.
e user's activity collected by the ARM was executed as the candidate item set in the recommended module: Besides, we assume that three factors affect the activity rating.
ey are activity, user, and context. e 3D cube stores the activity rating, like the one shown in Figure 2. is cube is being named Selective Cluster Cube (SCC). e Selective Cluster Matrix (SCM) represents each candidate activity with different contexts. In a specific context, a candidate item is expressed as an array. Every element of the array is the selective probability S w (P u , c j ) that the user "u" has chosen the activity w under the situation "c j ." In this paper, the selective probability was calculated from activity history. Users belonging to the same group consist of the same candidate activity Set A s, and the users have shared a standard SCC.
For example, in Figure 2(a), "skipping" was the RA (item). e equivalent context c 0 was [non-office hour, home, comfortable weather]. e probability of user 1 for "skipping" under the context of the non-office hour, home, and comfortable temperature was 0.76. In the same way, the probability of user 2 for "skipping" was 0.64. If the participant is idle, the system consecutively detects the present context and asks all the activity arrays that match SCC's current context. For example, in Figure 2(b), if the system tracks situation c0, activity arrays with context C 0 will be obtained from the SCC. e rating of candidate activity items was done according to a grade function, gTotal(S w ), and it measures the combined probabilities of the target user and the other users in the same group within the Time Window (T w ). is is shown as follows: where S w (P u , c j ) indicates the probability of the target user "u" perform activity "w" under context j, and S w (P userset , c j ) is the probability of the other users in the same group perform item "k" under the same context j. If "n" stands for the number of users in the user group, then S w (P userset , c j ) is calculated as follows: π is the influence parameter and is defined as At last, the item list was recommended to the user. e system also recorded the history of user interaction with the items. e participants' historical data consisted of the present context, and the things are opted by the participant from the recommendation list.

Proposed Security and reat Model.
e following are the description of the security and threat model: (i) For starters, smartphone users, RSPs, and PPAs do not discuss individual benefits and reputations. Furthermore, they execute each task and performance characteristics following the set design. (ii) Secondly, the communication of mobile users with RSP using anonymous identities is done candidly. We believe that protected channels (e.g., SSL protocol or other encryption protocols) are used among system entities in the medical data communication process.

Security and Communication Networks
(iii) irdly, privacy leakage is the biggest threat for mobile users (e.g., disclosing personal data of high sensitivity and mining personal privacy through data analytics) when the formalized data about AP, UT, PA, and other similar data are transmitted to RSP. Yet, giving normalized data, to some extent, will protect user privacy. Nonetheless, user's private information (e.g., similarities of other's favorites) is also a botheration for mobile users when an application recommendation is requested for themselves. (iv) Fourthly, the semitrusted nature of RSP can make it achieve the functionalities based on system design. But simultaneously, it is also conscious of user privacy, and there are possibilities of user privacy leakage when it obtains any helpful information. (v) Lastly, PPA, which is also semitrusted, can complete allocated system tasks according to system design through PPA that is also conscious about protecting user privacy.

Encryption Schemes.
is section consists of an introduction briefing HE and an application used in our schemes, followed by summarizing notation for easy reference. Finally, a description of the detailed protocol of the two proposed methods is given correspondingly.

Homomorphic Encryption.
Homomorphic Encryption (HE) permits computations to be performed over cipher texts. e decryption results are identical as if the HE operations had been carried out on the Plaintexts (PT) [34][35][36][37]. As an illustrative example, assume two plain texts Txt 1 and Txt 2, and their equivalent Ciphertexts (CT) txt 1 ←H M enc (txt 1 , key), txt 2 ←HM enc (txt 2 , key). An encryption scheme is additively homomorphic if txt 1 + txt 2 � HM dec (txt 1 ⊞ txt 2 , key)/multiplicatively homomorphic if txt 1 × txt 2 � HM dec (txt 1 ⊠txt 2 , key) where ⊞ and ⊠ stand for the homomorphic addition and multiplication operations, correspondingly. A few HE schemes are either additively homomorphic/multiplicatively homomorphic, like [28]. e schemes that come into this category are said to be Partially Homomorphic Encryption (PSE). e schemes that support both additions and multiplications but are restricted only a few times are called Some What Homomorphic Encryption (SWHE).

PRIPRO Recommendation System.
is is an interactive procedure that is inclusive of the recommendation requestor, RSP, and PPA. As a first step, a request is sent to RSP by the recommendation requestor, and in turn, the PPA sends back a set of protected user dataset to the requestor on a condition [39] that the user validity check is positive. After that, the requestor processes user data embracing HE, and the processing result is sent to the RSP. Next, RSP does HEs on the processing result, which is necessary for producing recommendations for preserving privacy. Finally, the requestor creates final recommendations on his MD using the calculated results sent from RSP. In this way, HE user data are processed for making final decisions [40,41].
A recommendation requestor, e.g., user p, sends his request Mrk p (Uid p ), Mrk PPA (Mrk p (Uid p )) to SP, where Mrk p (Uid p ) is the signature of p on his unknown identity and Mrk PPA (Mrk p (Uid p )) is PPA signature on p's signature. RSP verifies the validity of p with the support of PPA. If the check is positive, RSP computes the correlation of the user p with the rest of the users (user k) Crl(p, j) based on Equation (5); otherwise, RSP's performance is null.
where F p i′ (AP) User p's formalized AP value is related to application i', and the rest of the symbols imitate the same representation style. e correlations are preserved by Enc(c 1 ) at context x (the concrete i of Enc(c i ) is determined based on which context it is). en, RSP encryptsCrl s (p, j) as Enc(Crl p,j s ) � Enc p(K pub ), Crl s (p, j) with user p's public key p(K pub ) and returns a set of encrypted user's correlations c, Enc(Crl p,j s ), p ≠ j , to user j. After getting c, Enc(Crl p,j s ), p ≠ j , user j decrypts Enc(Crl p,j s ) with his secret key j(K sec ). After that, it chooses Enc(s 1 ) following time context c and obtains a set of user's accurate correlations Crl(p, j), p ≠ j by removingEnc(s w ). It is to be noted that user "j" is anonymous of anyone correlative user's true identity, user "p," for instance, because none of the User's ID is known. en, user "j" encrypts Crl(p, j) intoHM enc j(K HM ), Crl(p, j) with his HE key j(K HM ) and transmits a set of encrypted values HM enc j(K HM ), Crl (p, j)}, p ≠ jto RSP. Consecutively, RSP sums up the encrypted user correlation based upon HE computations: by making use of homomorphic operation PaillierMul, refer to equation (6): HM enc j K HM , Crl(p, j) .
en, RSP sums up the formalized values of AP, UT, and PA with user correlation by using homomorphic operation PaillierExp, refer to equation (7): RSP further sums up the encrypted cumulative result by using PaillierMul again (i.e., equations (8) and (9)): Security and Communication Networks is encrypted sum will be sent to the user's MD to be used to compute a selection of features. Algorithm 1 explains those as mentioned above serial HE activities on encrypted user correlations.

Dataset.
In recent times, datasets for distinct domain and context recognition have public accessibility. A realtime laboratory setup is done for the MIT Place Lab dataset. MD with integrated home settings was used to record the volunteers, and cameras fixed all over the houses were used for the annotation. In [42], the trial dataset was executed in a home system to monitor activity. is dataset utilization is effortless for binary sensing nodes (reed switches, pressure mats, etc.) and primarily comprises month-long readings. e dataset's existence in [43] spotlights individuals' day-today actions, recording a person's everyday life over 16 days. e Opportunity dataset [44] gives a tremendous recording of 12 subjects doing activities at dawn (activities of daily living (ADL)) in a room attached to a kitchen. is dataset allows the use of sensors installed in the respondents' bodies and the environment, and it consists of over 25 hours of sensor information. At last, the TUM Restaurant database was collected and shared with the public for study purposes in fields such as markerless human performance capture, motion edge detection, and human activity recognition. Video data from 4-Cameras, Radio-Frequency Identification (RFID) tags, reed switch test results, and activity tags were included in the data source. Subsequently, the researcher [45] presented a standard dataset for estimating sensor migration in activity recognition. e data contains 33 PA identified using 9 inertial sensor units from 17 subjects.
(1) PAMAP Dataset. As part of the aerobic activity, monitoring uses specific instances, and the PAMAP recording was performed using a system developed in the previous Physical Activity Monitoring for Aging People (PAMAP) design [46]. e information was analyzed in August 2020. At the time of data collection, wired 3D-IMUs and an HR monitor were used as sensing devices, and a Sony Vaio Ultra-Mobile PC , p ≠ j, the set of data vectors of other users of application "i" at context "c," secured by Enc(s 1 ) Output: {M, N}: the data to calculate final recommendations (1) While p ≠ j Do (2) For Each activity "a" performed by "p" and "j," Do (3) Measure M: the resultant value from the equations below steps (4) and (5) (4) Measure the value as HM enc j(K HM ), Enc(s w ) * (5) Measure N: the resultant value as HM enc j(K HM ),  Security and Communication Networks (UMPC) was used a retrieve data from things. Each of the 8 test subjects implements a one-hour-long data collection task. us, the data collection comprises 8 hours of data roughly. e dataset is open access and can be downloaded from the research work website1 for research purposes.
(2) Data Set for Monitoring Physical Activity in PAMAP2.
e PAMAP2 Physical Activity Monitoring datasets contain 18 comprehensive physical activities (such as stepping, riding a bicycle, and sporting events) undertaken by 9 subjects while wearing three inertial measurement systems and a heart rate sensor. More than 10 Hrs. of the survey was conducted in total, with approximately 8 hours labelled as one of the 18 activities. e data is accessible to the public and can be retrieved from the PAMAP research work's website2. Furthermore, the type of data labelled "PAMAP2 Physical Activity Monitoring Data Set" [47] is included in the UCI Machine Learning repository [48]. It measures by using 3 Colibri wireless inertial measurement units that consist of a timestamp, IMU hand, chest, and ankle. It contains temperature, 3D acceleration data, gyroscope data measured in rad/s, and magnetometer data (Î¼T).
(c) Real World (HAR). e dataset contains motion, location services, gyroscope, light, magnetic field, and environmental noise data from fifteen subjects age (31.9 ± 12.4⟶height; 173.1 ± 6.9⟶weight 74.1 ± 13.8; 8 Males and 7 Females) achieving behavior such as hopping, resting, standing, sitting, running/jogging, and walking. For each exercise, the rapid action of the chest, forearm, head, chin, thigh, shoulder blade, and waist was tracked in parallel. Each activity lasted approximately 10 minutes, excluding jumping (1.7 minutes) due to physical exercise.

Experimental Evaluation
Our approach's evaluation is done experimentally utilizing 3 real-time datasets and extensively used metrics with diverse parameters in this section. e experimental tests were carried out on a Desktop PC with a 2.9 GHz, 2.8 GHz, 2.2 GHz, Octa-Core CPU and 12 GB RAM, and a Galaxy S21 5G Android MD.

Accuracy Measures.
e accuracy of the proposed method's generated recommendations is considered by using the Mean Absolute Error (MAE) and the Root Mean Square Error (RMSE) [31,49]. Accuracy metrics were commonly used to assess the recommendation system. e MAE standard is defined in equation (10), with p i representing the assumed score and r i defining the overall score in the assessment. MAE is also used to measure the difference between the predicted and accurate scores. It should be acknowledged that lower values indicate a better prediction of guidelines. Equation (11), for instance, demonstrates RMSE [50,51]. RMSE is the same as MAE but with squared values. As an outcome, lower values are recommendable to RMSE.
4.2. Results. Figures 3 and 4 demonstrate the MAE and RMSE values obtained from the 3 datasets. However, all the datasets show an initial deviation by the model since the K-value increases, and all the three models follow a similar pattern.
Top-K Mining ( Figure 5). A metric that shows how well a user's favorite activities are protected. is metric is defined as TKM � 1/n n k�1 TPR k , where TPR k is the True Positive Rate (TPR) of the Top-K activities suggested by the data.

Robustness.
e sturdiness of the 2 schemes was tested, taking into account the following attacks: • On-Off Attack e impact of each attack over the dataset was calculated using the accuracy parameter, as shown in (1) Bad-Mouth Attack. e results of 3 datasets of bad-mouth attack simulation are shown in Figures 6(a)-6(c). e 2 dataset's results are the same, and bad-mouthing attackers impact the 3 rd dataset, Real World, but the impact has not increased with time flying. e technical details like misbehaving protocol like bad guy's bad breathing, crime attack, or escape sequences are included to identify vulnerabilities, defenses, and functionality. us, the outcome of the proposed method shows that it can resist bad-mouthing attacks to some extent. e on-off attack in PAMAP has a specific impact on recommendation accuracy initially. However, the impact reduces with time flying due to increased data volume for constructing the regular database in different time slots. e impact of the on-off attack on accuracy recommendation in PAMAP-2 and real-world oscillates with time flying, yet the oscillation range is quite admissible. e oscillation happens due to the database construction process that performs only at the time of generating recommendations.

User Interface of ARS with Implementation.
e main goal of the experimental tests is to evaluate that the proposed mobile recommendation current study is better, practical, and novel. Evaluation tools are classified into two main Security and Communication Networks categories: computational space and computational time.
e processing time needed for DDDMS is a critical factor in determining the results of the developed framework.
Processing time and the quality of DDDMS can both influence system efficiency. Offline and online data processing methods are used based on user requirements to reduce      computational time and make personalized decisions. Several works have aimed to go beyond classical accuracy metrics in the monitoring and development of RS design, focusing on recommendation diversification as a method of improving user serviceability. Because, in skipping alone, the user can attain breathing problems while doing exercise, so, here, we primarily consider understanding skipping alone. ere will be some disturbance while calculating skipping in the user. is will be taken into consideration for RS. An ARS named "Diversify" was developed using the PRIPRO framework, illustrated in Figure 8, which includes both the GUI and deployment.

Conclusion
PRIPRO, a novel privacy security system for personalized fitness reviews, was introduced in this paper. Besides providing information on the device structure and module design, we quantitatively conduct an extensive experiments model against various datasets. We concluded that the proposed PRIPRO model is ideally suited to prescribe physical activity to the application user by considering the current context and maintaining the privacy of the user's data. PRIPRO is comprised of HE schemes in securing user data while making preferred decisions. PRIPRO model uses only arithmetic operations that are precisely accurate with HE methods. It can also make DDDMS using a pretrained model, while the client's data is not in the server training collection. e tests show that PRIPRO can provide high bandwidth recommendation solutions while still improving the state-of-the-art performance.

Data Availability
No data were used to support this study.

Conflicts of Interest
e authors declare that they have no conflicts of interest.