Source-linked AI summary
Trust Evaluation Mechanism for User Recruitment in Mobile Crowd-Sensing in the Internet of Things
Nguyen Binh Truong, Gyu Myoung Lee, Tai-Won Um, Michael Mackay
TL;DR
MCS trust assessment is challenged because QoD alone may not capture service satisfaction and many users lack prior Experience with one another. The paper proposes the E-R model and trust-based recruitment using virtual interactions, Experience, and Reputation; the scheme improves QoS and detects malicious users.
Problem
QoD scores may not fully reflect consumer satisfaction, while many MCS users have no prior Experience with one another.
Method
The E-R model quantifies virtual interactions from contributed-data quality to calculate Experience and Reputation for trust-based user recruitment.
Results
The Trust-based scheme generally achieves better QoS, detects malicious users, and identifies 80% of malicious users among the 10% with lowest reputation values.
Takeaways & Limitations
Trust relationships based on Experience and Reputation support selecting trustworthy contributors and maintaining better service quality in untrustworthy environments.
Abstract
from arXiv · showhide
Mobile Crowd-Sensing (MCS) has appeared as a prospective solution for large-scale data collection, leveraging built-in sensors and social applications in mobile devices that enables a variety of Internet of Things (IoT) services. However, the human involvement in MCS results in a high possibility for unintentionally contributing corrupted and falsified data or intentionally spreading disinformation for malevolent purposes, consequently undermining IoT services. Therefore, recruiting trustworthy contributors plays a crucial role in collecting high-quality data and providing better quality of services while minimizing the vulnerabilities and risks to MCS systems. In this article, a novel trust model called Experience-Reputation (E-R) is proposed for evaluating trust relationships between any two mobile device users in a MCS platform. To enable the E-R model, virtual interactions among the users are manipulated by considering an assessment of the quality of contributed data from such users. Based on these interactions, two indicators of trust called Experience and Reputation are calculated accordingly. By incorporating the Experience and Reputation trust indicators (TIs), trust relationships between the users are established, evaluated and maintained. Based on these trust relationships, a novel trust-based recruitment scheme is carried out for selecting the most trustworthy MCS users to contribute to data sensing tasks. In order to evaluate the performance and effectiveness of the proposed trust-based mechanism as well as the E-R trust model, we deploy several recruitment schemes in a MCS testbed which consists of both normal and malicious users. The results highlight the strength of the trust-based scheme as it delivers better quality for MCS services while being able to detect malicious users.
I. INTRODUCTION
MCS extends IoT data collection through mobile devices but faces data-quality and security risks from human participation. The paper proposes E-R trust evaluation and trust-based recruitment to select reliable contributors and improve service quality.
- Background: MCS leverages built-in sensors and applications in mobile devices to support diverse IoT sensing campaigns.
- Problem: Low-quality, corrupted, or falsified data can undermine MCS services and increase system vulnerabilities, risks, and potential attacks.
- Approach: The E-R mechanism evaluates trust relationships between MCS users using Experience and Reputation indicators derived from QoD-assessed virtual interactions.
- Approach: The proposed recruitment scheme selects trustworthy contributors based on trust relationships between service requesters and potential participants.
- Evaluation: The evaluation combines a QoS model based on QoD assessment with simulations comparing trust-based and other recruitment mechanisms in an MCS testbed.
B. Related Work
Prior MCS recruitment research addresses cost, coverage, QoI, reputation, and trust, but important approaches assume trustworthy users or require task-similarity analysis. This paper targets a practically deployable trust mechanism for recruiting trustworthy users.
- Existing recruitment objectives: Existing recruitment schemes commonly optimize sensing cost, energy efficiency, coverage, or user location while maintaining task requirements.
- Quality-aware recruitment: QoI- and QoD-aware approaches consider task requirements or participation history, but some assume recruited users provide compliant data or require similarity features among tasks.
- Reputation and trust: Reputation-based methods address malicious users through statistical readings or community voting based on shared sensing-task participation.
- Research direction: The paper proposes a trust evaluation mechanism intended to recruit trustworthy users while remaining practically deployable for real-world services.
- System setting: The proposed mechanism is deployed on a centralized MCS platform using an indirect, participatory sensing model.
B. E-R Trust Mechanism in the MCS Platform
The E-R mechanism simplifies the REK trust model by using Experience and Reputation to evaluate relationships between MCS users. QoD assessment and feedback support trust-based recruitment, while Experience is updated through interactions and QoD thresholds.
- Trust predicts whether a mobile user will provide high-quality data for a requested MCS service.
- The REK model includes Reputation, Experience, and Knowledge, with Knowledge representing direct trust and the other indicators representing indirect trust.
- The E-R model omits Knowledge because its required information raises privacy and practical retrieval concerns, relying on Experience and Reputation.
- QoD measures how contributors fulfill sensing tasks, while feedback can complement QoD because scores do not fully represent consumer satisfaction.
- Experience is an asymmetric relationship updated after interactions to reflect how much a trustor trusts a trustee.
- Experience increases after cooperative interactions, decreases after uncooperative interactions, and decays during absent or neutral interactions.
B. Analysis and Discussion for Experience Model
The Experience model keeps trust values bounded while making updates depend on QoD, current Experience, and interaction type. Its decay design reduces strong relationships more slowly, while repeated cooperative interactions approach the maximum.
- Experience values are normalized to the range (0, 1), with maxExp = 1 and minExp = 0.
- The increase function remains below 1 and approaches 1 asymptotically under repeated cooperative interactions.
- 0 < Expt < 1 for all t under the model’s stated parameter conditions.
- The decrease rate β should exceed 1 because strong relationships are harder to gain but easier to lose.
- The decay model ensures that even strong relationships decline, while strong relationships decay more slowly than weak ones.
C. Reputation Model
Reputation represents the community’s overall opinion of a user and is especially useful when a requester and provider lack direct Experience. The model derives it from the Experience relationships of other users.
- Reputation is vital when a service requester and data provider have no prior Experience relationship.
- Reputation aggregates the opinions of users who have prior Experience with the evaluated user.
- The model classifies Experience relationships as Positive or Negative using a predefined threshold θ.
- Overall Reputation combines separate positive and negative reputation components.
D. Mathematical Analysis for Reputation Model
The reputation model resolves correlations among users’ reputations by formulating reputation propagation as a Markov-chain problem. Its resulting overall reputation vector exists and is unique.
- The proposed reputation vector is recursively calculated from other users’ reputations and corresponding Experience relationships, creating correlations among all users.
- The positive reputation vector is represented through a transition matrix whose entries depend on Experience relationships and a threshold condition.
- The resulting Markov chain is strongly connected because it incorporates random jumps, making the positive reputation stationary distribution unique.
- The negative reputation vector also exists uniquely, so the overall reputation vector exists and is unique.
E. Final Trust Value
Final trust combines Experience and Reputation, while the evaluation models user quality through Beta-distributed QoD patterns. The testbed includes normal, low-quality, and intelligent malicious user behaviors.
- Final trust is calculated as a weighted sum of Experience and Reputation values, with positive weights that sum to 1.The weighting factors can be tuned autonomously using machine learning or semantic reasoning.
- The testbed compares trust-based recruitment with Average and Polynomial Regression predictive models.
- User models: QoD scores from traffic and parking sensors in Santander fit the Beta probability-distribution family after normalization to (0, 1).
- User models: High-quality users concentrate QoD scores around 0.75–0.85, whereas low-quality users mostly produce scores around 0.5–0.65.The high-quality and low-quality groups are modeled with distinct unimodal Beta distributions.
- User models: Intelligent malicious users are modeled as producing very high QoD before recruitment and intentionally very low-quality data after recruitment.Their QoD behavior follows a bimodal Beta distribution.
- User models: The malicious-user model uses γ = 0.7, representing high-quality behavior in 70% of sensing tasks and very low-quality behavior in 30%.
B. QoS Evaluation Model for MCS Services
The QoS evaluation model measures service quality from the QoD of sensing tasks and their contributing participants. It reflects how low-quality data affects MCS service operation and cost.
- Low-quality data lowers system efficiency, misleads operations, increases overhead and cost, and imposes vulnerabilities and risks.
- A service request consists of sensing tasks, each fulfilled by participants providing datasets with associated QoD scores.
- The QoS of a service request is proportional to the product of the natural logarithms of its sensing-task QoD scores.
- Each sensing-task QoD score is calculated as the average QoD of its participants’ datasets.Participants in the same task are normally required to collect the same type of data, creating redundancy.
C. Trust-based, Average, and Polynomial Regression User Recruitment Schemes
Three recruitment schemes use prior sensing-task QoD information to select users and produce QoS scores for requested services. The trust-based scheme instead uses maintained trust relationships with service requesters.
- All three recruitment schemes aim to recruit users expected to provide high QoS scores for sensing tasks.
- The schemes rely on QoD scores from users recruited in previous sensing tasks and share the same inputs and QoS-score outputs.
- The comparison includes Trust-based, Average, and Polynomial Regression recruitment schemes.
- Trust-based User Recruitment scheme: Trust-based recruitment establishes and maintains Experience–Reputation trust relationships, then recruits users with the highest trust values to a requester.
- Trust-based User Recruitment scheme: The trust-based algorithm initializes Experience, Reputation, and Trust records, recruits participants for each task, calculates QoD, and updates Experience afterward.
13 Return out
The paper compares average-based and polynomial-regression-based QoD recruitment schemes. Both schemes select users using QoD histories, evaluate collected data, and update service outputs and stored scores.
- Average-QoD User Recruitment scheme: Average-based recruitment selects participants with the highest average QoD scores from a maintained user list.The algorithm initializes an AVG matrix, recruits users for each sensing task, computes QoD for collected data, and updates AVG.
- Polynomial Regression-based QoD User Recruitment scheme: Polynomial regression-based recruitment predicts each user’s next QoD score from historical contributions and recruits users with the highest predictions.A 3-degree polynomial fitted by least squares provides the predictive model.
- Polynomial Regression-based QoD User Recruitment scheme: The polynomial scheme computes QoD for collected data and updates its QoDScore history after each sensing task.The algorithm stores previous QoD scores and updates the requested service output after processing tasks.
13 Return out
The evaluation uses a Matlab testbed with low-quality, high-quality, and malicious users and compares recruitment schemes under shared inputs. The Experience model uses QoD thresholds and tunable interaction parameters, while reputation is computed iteratively for scalability.
- Testbed simulation scenarios: The Matlab testbed includes low-quality, high-quality, and malicious users and compares three recruitment schemes using identical inputs and QoS-score outputs.The scenarios vary service requests, sensing tasks, participating users, and the proportion of malicious users.
- Experience Model Parameters: QoD scores ≥0.6 represent cooperative interactions, whereas scores ≤0.3 represent uncooperative interactions in the Experience model.The bootstrap Experience value is 0.3, with maxExp=1 and minExp=0.
- Experience Model Parameters: α=0.1 controls the maximum Experience increase, so smaller α values require more interactions to build a strong relationship.The Decrease and Decay models use separately tuned parameters δ and γ to produce reasonable curves.
- Reputation Model: The iterative Reputation calculation converges in 25 to 32 iterations for 200 to 1000 users with damping factor 0.85 and scales roughly linearly in logarithm of N.The iterative method is selected instead of the algebraic method because the latter requires roughly N^3 operations.
C. Results and Discussion
The trust-based recruitment scheme generally delivers higher QoS than alternative schemes, particularly as malicious users increase, while the E-R model identifies users associated with low-quality or malicious behavior. Its benefits are balanced by lower-resource alternatives in low-malicious-user settings and by future needs for context-specific adaptation.
- Service-volume evaluation: After about 15 requests, the Trust-based scheme stabilizes and achieves QoS scores of about 3.35 to 3.55, exceeding Average-based and Regression schemes.Average-based takes about 35 requests and Polynomial Regression about 70 requests to reach consistent QoS.
- Malicious-user detection: The E-R model heavily penalizes users with occasional very low QoD scores, rapidly lowering their trust relationship and reputation value.After 20 service requests with 10% malicious users, 80% of malicious users appeared among the 10% of users with the lowest reputation values.
- Malicious-user prevalence: As the malicious-user percentage rises from 0% to 25%, QoS decreases across recruitment schemes, while increasing requested services generally improves QoS except under Random Selection.At 15% malicious users, Trust-based QoS rises from about 3.2 after 10 services to about 3.6 after 160 services.
- Malicious-user prevalence: With 160 requested services, the QoS gap between Trust-based and Regression expands from 0.07 at 10% malicious users to 0.18 at 25%.The corresponding QoS pairs are 3.65 and 3.58, then 3.49 and 3.31.
- Trade-offs and future work: When malicious users comprise less than 10%, Average-based offers similar QoS with fewer computing resources, while Reputation should be executed periodically rather than at every evaluation.The paper also identifies context-aware parameter adaptation and customization for specific MCS use cases as future directions.
- Mechanism and deployment: The proposed E-R mechanism combines QoD-based virtual interactions with Experience and Reputation to support trust-based recruitment in MCS.The authors report better QoS in most cases and practical implementation in real-world IoT services through the Wise-IoT project.