Trust and Reputation Management Systems

• Trust and Reputation Management Systems (TRMSs) are socio-technical frameworks that collect and aggregate atomic trust signals to compute measurable scores for decision-making across diverse domains.
• They employ heterogeneous evidence—such as direct interactions, recommendations, and public ratings—combined through statistical, graph-based, and machine learning models to infer trust dynamics.
• Architectural designs vary from centralized to decentralized systems, often integrating TRMS into access control and consensus protocols to mitigate threats like Sybil attacks and bad-mouthing.

Trust and Reputation Management Systems (TRMSs) are socio-technical systems that collect atomic trust signals, aggregate them into trust and reputation scores, and expose these scores to support selection, access, pricing, and governance decisions. In the current literature, trust is a subjective belief of a trustor about a trustee’s expected future action, trustworthiness is the inherent, objective quality that makes an entity deserving of trust, and reputation is the collective perception of that trustworthiness derived from publicly accessible and aggregated information. Across wireless self-organizing networks, IoT, supply chains, data sharing, 6G, and distributed consensus, TRMSs are used to evaluate both entities and, increasingly, data and compliance behaviour over time (Wu et al., 19 Aug 2025, Putra et al., 2022).

1. Core concepts and trust objects

A standard conceptual distinction separates trust, reputation, and trustworthiness. Trust is “a subjective belief held by one entity (the trustor) about another (the trustee), representing the perceived probability that the trustee will perform a specific future action as expected, even without direct monitoring.” Trustworthiness is “the inherent, objective quality of an entity that makes it deserving of trust,” while reputation is “the collective perception of an entity’s trustworthiness, derived from publicly accessible and aggregated information such as ratings, reviews, and referrals” (Wu et al., 19 Aug 2025). In 6G networking, the same distinction is stated more operationally: trust is a subjective expectation gradually built from repeated interactions, reputation is an aggregated trust degree from multiple participants over a temporal horizon, and TRM is the mechanism that quantitatively assesses the trustworthiness of network participants from direct experience or recommendations (Putra et al., 2022).

TRMSs also differ in what they evaluate. The data-sharing literature distinguishes data-centric trustworthiness from entity-centric trustworthiness. Data-centric trustworthiness concerns intrinsic quality dimensions such as authenticity, integrity, validity, accuracy, completeness, consistency, timeliness, relevance, precision, traceability, and uniqueness. Entity-centric trustworthiness concerns behavioural and organisational properties of data providers, data consumers, intermediaries, and trustees, including compliance, accountability, security practices, transparency, consent handling, and historical reputation (Wu et al., 19 Aug 2025). This distinction reappears in supply-chain systems that separate product-level and participant-level trust. TrustChain maintains product/lot-level quality ratings and seller-level trust per commodity type, while DeTRM separates commodity trust from behaviour-based participant trust and endorsement-based participant trust (Malik et al., 2019, Putra et al., 2022).

The literature also distinguishes local trust from global reputation. In decentralized IoT access control, local trust is a subjective, pairwise quantity computed by each service provider about each service consumer, whereas global reputation is an objective, system-wide view computed on-chain from all interactions between that consumer and multiple providers (Putra et al., 2019). A similar distinction appears in online and marketplace settings, where direct interactions generate individual trust, while ratings, reviews, and other propagated signals generate reputation. This suggests that TRMS design begins by deciding whether the target of evaluation is a node, a service, a dataset, a product lot, a principal, or a role-specific manifestation of any of these.

2. Evidence, inference, and formalization

TRMSs rely on heterogeneous evidence. The literature repeatedly groups trust signals into direct observations, recommendations or referrals, and public or system-level information. In 6G resource sharing, direct interaction evidence includes fulfilment of obligations, SLA violations, interruptions, and “undesirable incidents,” while feedback from resource users supplies recommendation-like evidence. In data-sharing systems, explicit ratings coexist with implicit monitoring signals such as QoD measurements, compliance checks, SLA events, and security logs (Putra et al., 2022, Wu et al., 19 Aug 2025). In IoT surveys, the same structure appears as direct trust, indirect trust, and reputation propagation, frequently combined through weighted aggregation, belief combination, or probabilistic updates (Aaqib et al., 2023).

The computational models span statistical, rule-based, graph-based, and learning-based families. The survey literature lists time-decay weighted averages, sliding windows, Bayesian updates with Beta priors, Dempster–Shafer Theory, fuzzy logic, PageRank- or EigenTrust-style graph propagation, and ML- or GNN-based estimators as standard mechanisms for turning atomic evidence into longer-term trust or reputation (Wu et al., 19 Aug 2025). In online communities, the Dynamic Interaction-based Reputation Model computes interaction value as $I_n = I_{b_n} + I_{c_n}$, where the cumulative part is $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$, and updates trust by $T_n = T_{n-1} * \beta^{\Delta_n} + I_n$. This model uses only implicit temporal information and was shown to mimic the behaviour of Reddit and MathOverflow reputation systems using temporal frequency of interaction alone (Yashkina et al., 2019).

A different line of work treats compliance as a first-class trust signal. The Automated Compliance Engine formalizes legal and organisational policies into obligation-centric Prolog rules, audits immutable event logs, emits typed violations, and computes a per-principal compliance score $Comp(p, W_k)\in[0,1]$. Its quantitative core uses Volume, Duration, Breadth, and rule Criticality, with normalized severity $S(M) = 1 - e^{-(M / \alpha)}$, time-decaying penalties $$\text{Penalty}_k = (\text{Penalty}_{k-1} \cdot e^{-\lambda}) + Sev_{p, W_k}$$, and the compliance score $Comp(p, W_k) = 1 - \tanh(\text{Penalty}_k)$ (Wu et al., 3 Jan 2026). This moves TRMS computation away from subjective ratings and narrow QoS metrics toward a white-box, regulation-aligned trust metric.

Formalization has also been pursued at the system semantics level. A process-algebraic framework for collective adaptive systems models trust and reputation by embedding trust values into agent behaviour, supports trust-based communication and dynamic communities, and yields a trust labeled transition system amenable to model checking with a dedicated trust temporal logic. In that framework, placeholders for unrated interactions and explicit feedback actions become part of the operational semantics, allowing bad-mouthing, ballot-stuffing, collusion, on-off behaviour, Sybil behaviour, and white-washing to be represented as process behaviour rather than only as simulation scenarios (Aldini, 2016).

3. Architectural patterns and implementation

A general architectural taxonomy distinguishes centralized, decentralized, and federated TRMSs. Centralized systems rely on a single Trusted Third Party that gathers feedback and computes trust scores; decentralized systems distribute trust operations across nodes, often with DLT or blockchain; federated systems combine local trust computation with periodic upload of local trust vectors to an aggregator that computes a global reputation view (Wu et al., 19 Aug 2025). The same taxonomic line reappears in IoT surveys, which split systems into traditional trust-management-based models and AI-based models, then further into decision models, evaluation models, data-mining models, and deep-learning models (Aaqib et al., 2023).

Blockchain-based TRMSs typically separate on-chain evidence from off-chain actions. In the 6G resource-sharing architecture, a consortium blockchain includes a TRM contract, a Resource Manager contract, and per-agreement Sharing Agreement contracts. Resource listings, bids, allocation decisions, creation and termination of agreements, SLA-relevant events, feedback, and trust scores are on-chain; actual computation and data transfers are off-chain (Putra et al., 2022). TrustChain uses a three-layer architecture—Data layer, Blockchain layer, Application layer—in which raw IoT data remain off-chain while hashes, ratings, and aggregated trust values are committed on-chain (Malik et al., 2019). DeTRM extends the same pattern with a four-layer structure $$\mathbb{SC} = \langle \mathbb{L}_{app}, \mathbb{L}_{bc}, \mathbb{L}_{data}, \mathbb{L}_{phy} \rangle$$, where commodity contracts and trust-and-reputation contracts operationalize create, produce, monitor, inspect, and trade primitives on a consortium blockchain (Putra et al., 2022).

Some architectures integrate trust directly into authorization or consensus. In decentralized IoT access control, a private Ethereum blockchain hosts an Attribute Provider contract, a Trust and Reputation System contract, and a Policy contract, so that trust and reputation become explicit attributes in ABAC policies and token issuance (Putra et al., 2019). In TRM-Raft, a Blockchain-based Trust and Reputation Model is embedded into Raft leader election and log replication. Reputation-aware election excludes low-reputation nodes from leadership, while Schnorr-signature verification of replicated logs feeds objective tampering evidence into the trust layer (Zhang et al., 9 Jul 2026). A plausible implication is that TRMSs increasingly function not as external scoring services but as control-plane mechanisms that directly determine role eligibility, access rights, and coordination authority.

Architectural choices also reflect device capabilities. The DLT-backed IoT reputation framework that combines Tangle and Blockchain introduces a WeakReq protocol so that low-power devices can issue requests without doing PoW or storing full ledger data, while miners attach their messages to the Tangle and blockchain incentives maintain the hash rate needed for security (Mirhosseini et al., 2021). This design makes explicit a recurrent TRMS engineering theme: evidence collection, trust computation, and decision enforcement are often decomposed across heterogeneous participants with very different computational budgets.

4. Threat models, robustness, and privacy

The attack surface of TRMSs is well established. Across 6G, IoT, and data-sharing work, the canonical threats include self-promotion, bad-mouthing, ballot-stuffing, Sybil attacks, whitewashing or newcomer attacks, on-off attacks, collusion, privacy leakage, and cold start (Putra et al., 2022, Wu et al., 19 Aug 2025). Classical e-market analyses add reciprocity, retaliation, reputation-lag, value-imbalance, sudden-exit, and multiple-identity attacks, emphasizing that reputation manipulation and economic exploitation are intertwined (Gaur et al., 2013).

Mitigations fall into several families. Blockchain-based 6G TRM uses source-aware feedback and aggregation, evidence-backed negative feedback, escrow fees, real attributes that “cannot be easily changed,” and a global, strongly penalizing reputation to deter self-promotion, bad-mouthing, ballot-stuffing, Sybil, whitewashing, and on-off attacks (Putra et al., 2022). The hybrid Tangle-plus-Blockchain IoT framework couples payment and rating, requires valid paid trades for feedback, uses Proof-of-Burn and fees for weak-device requests, and leverages immutable identity histories to make slandering, ballot stuffing, and whitewashing economically costly (Mirhosseini et al., 2021). In e-commerce, dynamic weighting of individual versus shared reputation and value-sensitive reputation updates were proposed specifically to reduce the effect of ballot stuffing, bad-mouthing, reciprocity, retaliation, re-entry, and the Value-Imbalance Problem (Gaur et al., 2013).

Robustness can also be statistical rather than rule-based. COBRA encapsulates local interaction histories into models, then aggregates model outputs with a Context-aware Bernoulli Neural Network. In simulations with 51 malicious agents out of 100, malicious advisers share fake models that output $1-p$ instead of $p$, yet COBRA maintains accuracy above 85% in 90 out of 100 tested distributions of interaction probabilities, and the paper states that it is resistant to the 51-percent attack (Zeynalvand et al., 2019). In regulated settings, ACE avoids “gaming the system” by using additive penalties, event grouping, rule criticality, and time-decay-only recovery, so compliant actions do not directly average away severe violations (Wu et al., 3 Jan 2026).

Privacy-preserving techniques constitute a parallel dimension of TRMS design. The data-sharing survey groups them into identity protection, data protection, and computational privacy. Identity protection includes anonymity and pseudonymity via PKI; data protection includes encryption in transit and at rest and data minimisation; computational privacy includes Zero-Knowledge Proofs, Secure Multi-Party Computation, and homomorphic encryption (Wu et al., 19 Aug 2025). Some systems pursue privacy by abstraction rather than cryptographic privacy. COBRA shares local ML models instead of raw interaction histories, thereby preserving context while withholding the underlying records (Zeynalvand et al., 2019). This suggests that privacy in TRMSs is addressed either by constraining what evidence is revealed or by cryptographically transforming how trust is computed.

5. Application domains and representative designs

In e-commerce and agent-mediated markets, TRMSs are designed to cope with information asymmetry and strategic cheating. The Dynamic Reputation System for agent-mediated e-commerce defines overall reputation as $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$0, where $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$1 is individual reputation, $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$2 is shared reputation, and $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$3 increases with experience between a specific buyer and seller. It further introduces a value-sensitive update through $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$4, where $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$5 is transaction value, and discounts repeated buyer–seller interactions through $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$6, so the magnitude of reputation change grows with transaction value and shrinks with repeated pairwise exchanges (Gaur et al., 2013). A different e-commerce architecture introduces an “intelligent layer” that presents prefabricated feedbacks to a rater, updates the user’s trust degree in $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$7 according to like/dislike choices on feedbacks with known trustworthiness, assigns the same trust degree to the user’s own feedback, and then computes the product’s global trust score as a trust-weighted average of appreciations (Rahimi et al., 2014).

In IoT, trust often drives routing, access control, and service selection. Decentralized IoT access control computes local trust with an ageing function $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$8 and maps it to a bounded score with a Gompertz function $$I_{c_n} = I_{b_n} * \alpha * \left(1 - \frac{1}{A_n + 1}\right)$$9, while global reputation uses exponentially decayed interaction outcomes and a diversity factor $T_n = T_{n-1} * \beta^{\Delta_n} + I_n$0 (Putra et al., 2019). IoT surveys show that Bayesian, fuzzy, graph, entropy, SVM, Random Forest, ANN, Q-learning, DNN, RNN, and LSTM models have all been used to detect malicious nodes, estimate trust dynamically, or learn trust thresholds in social IoT, industrial IoT, vehicular networks, and smart environments (Aaqib et al., 2023).

Supply-chain TRMSs explicitly separate product quality from participant reliability. TrustChain computes event-level seller reputation as $T_n = T_{n-1} * \beta^{\Delta_n} + I_n$1, then aggregates it over time using a forgetting factor $T_n = T_{n-1} * \beta^{\Delta_n} + I_n$2 into $T_n = T_{n-1} * \beta^{\Delta_n} + I_n$3, and finally combines it with additional features $T_n = T_{n-1} * \beta^{\Delta_n} + I_n$4 into $T_n = T_{n-1} * \beta^{\Delta_n} + I_n$5 (Malik et al., 2019). DeTRM assigns commodity trust by $T_n = T_{n-1} * \beta^{\Delta_n} + I_n$6, where $T_n = T_{n-1} * \beta^{\Delta_n} + I_n$7 is derived from agreement with adjacent sensors, and defines participant reputation as a weighted combination of average commodity trust, behaviour-based trust, and endorsement-based trust (Putra et al., 2022). In both cases, trust becomes path-dependent and commodity-specific rather than a single seller-wide scalar.

In 6G and distributed coordination, TRMSs become infrastructure mechanisms. The 6G resource-sharing case study uses blockchain-based TRM to help resource users choose resource owners with sufficiently high reputation, with reputation updated recursively from SLA events and feedback (Putra et al., 2022). TRM-Raft integrates reputation into leader election and log replication, maintaining malicious leader ratio below 5% even with 40% Byzantine nodes while keeping throughput and latency close to vanilla Raft (Zhang et al., 9 Jul 2026). In data sharing, the survey literature argues that TRMSs must jointly evaluate data trustworthiness and entity trustworthiness across the phases of preparation and publication, data sharing transaction, and utilisation and governance, rather than reusing entity-only models from other domains (Wu et al., 19 Aug 2025). A plausible implication is that the application domain determines not only which signals are available, but also what the trusted object is and what decisions the resulting score will govern.

6. Evaluation practice, limitations, and research directions

Evaluation practice in TRMS research is heterogeneous. Surveys of IoT systems compare models through community-accepted performance metrics such as accuracy, scalability, delay, cooperativeness, efficiency, energy, context-awareness, dynamicity, recommendation support, and resilience to attacks (Aaqib et al., 2023). Data-sharing work calls for unified benchmarks, shared datasets, attack scenarios, and evaluation metrics for both data and entity trustworthiness, noting that most current systems are measured on synthetic data or domain-specific traces and that standard benchmarks are largely absent (Wu et al., 19 Aug 2025).

Some systems already report concrete performance and behavioural outcomes. ACE reports perfect recall on synthetic hospital datasets with 5% labeled violating entries, roughly linear runtime scaling, throughput around 75–100 entries/s on M1 Max, and memory below 500 MB even for 500k log entries (Wu et al., 3 Jan 2026). The 6G blockchain-based TRM simulation uses 500 Resource Owners, 600 Resource Users, and 100 iterations; with TRM, resource utilization increases significantly as the proportion of reliable resource owners grows, whereas without TRM utilization stays around 20% and convergence occurs at about 28 epochs (Putra et al., 2022). TRM-Raft reports that reputation-aware election and Schnorr-signature-based tampering detection keep malicious leader ratio below 5% even with 40% Byzantine nodes, with less than 10% throughput loss and less than 5% latency increase over vanilla Raft (Zhang et al., 9 Jul 2026). COBRA reports accuracy around 0.91 and RMSE around 0.26 on a real web-service dataset, outperforming HABIT, BRS, and TMSIoT, while also reducing overfitting relative to fully connected dense networks (Zeynalvand et al., 2019).

The literature also converges on a common set of limitations. Policy formalization still needs expert review in compliance-centric systems; synthetic evaluation remains prevalent; near-real-time trust computation may require optimisation; correlated violations can lead to over-penalization; and many frameworks assume immutable logs, honest authorities, secure identity management, or secure consensus without giving a full formal threat model (Wu et al., 3 Jan 2026, Malik et al., 2019). Blockchain-based systems repeatedly acknowledge de-anonymization risks, consensus overhead, the need for sharding or Layer-2 support, and the difficulty of transferring reputation across autonomous networks with different semantics (Putra et al., 2022).

Research directions are correspondingly consistent. The data-sharing survey identifies unified, context-aware evaluation frameworks, compliance as a trust metric, data-centric trust evaluation, accountability and incentive mechanisms, explainable trust evaluation, and Security-by-Design TRMSs as key priorities (Wu et al., 19 Aug 2025). IoT surveys emphasize multi-dimensional trust, resource-aware AI, ensemble and hybrid models, federated or edge-supported learning, and better resilience to sophisticated combined attacks (Aaqib et al., 2023). This suggests that the dominant trajectory of TRMS research is toward systems that are simultaneously context-aware, role-specific, bidirectional, privacy-preserving, explainable, and tightly integrated with the infrastructures they govern.