A blockchain-based framework for trust management and cache sharing reported higher detection accuracy than two named comparator schemes in a computer simulation, reaching 92.3% under the paper’s stated 10% noisy-feedback setting. Its reported latency was 0.49 seconds, compared with 0.58 seconds for Xu et al. and 0.47 seconds for Zhang et al.
The work is an arXiv preprint, not a report from a deployed network. The study used synthetic edge-server and mobile-user scenarios and included no human, animal or clinical participants. Its findings therefore describe how the proposed rules behaved under the tested simulation settings, rather than how real users or networks would fare.
What the model is designed to do
The proposed system is intended to identify trustworthy edge servers and mobile nodes, filter malicious nodes, encourage cache sharing and authenticate identities. It combines trust management with cache allocation and rewards, making the classification of participants central to the design.
Authentication begins with mobile users registering through a cloud registration center. They receive keys generated with RSA, a public-key encryption method, and then authenticate with a cloud authentication server.
The blockchain arrangement is designed around the limits of mobile devices. Only edge servers take part in consensus, and validation of each block is divided into two stages so that mobile users do not carry the full computational and energy burden.
For cache allocation, the framework uses max-min fairness, a rule intended to distribute resources while limiting extreme disparities between participants. Cooperative mobile users receive rewards based on the cache capacity they share and their trust score. The trust calculation also assigns credibility to each mobile-user evaluation and uses an indirect-trust weight of alpha = 0.9 when combining evaluations from multiple sources.
A small synthetic network produced the headline result
The main parameter-setting simulation used 10 edge servers and 20 mobile users. The server population was divided into 20% malicious servers, 20% low-quality servers and 60% high-quality servers. The scenario ran for 500 interactions and was repeated across 10 runs.
The researchers evaluated the model’s ability to distinguish the three server categories using accuracy, precision, recall, specificity and F1-score. These measures were calculated from a three-class confusion matrix and averaged across the categories, rather than treating one class as the only outcome of interest.
Several parameter sweeps were used to select settings for the trust calculations. Range 1 was selected for the trust-classification thresholds. At an alpha weight of 0.1, the reported specificity and recall were above 85%, while precision and overall accuracy were highest at that setting.
The trust-threshold weight beta was set to 0.8. At that setting, specificity and recall were reported above 0.8, with precision and accuracy reaching their highest reported values. The edge-server cache-status weight omega was set to 0.4, where specificity and recall were also above 0.8 and precision and accuracy were highest.
For cooperative mobile nodes, the cache-status weight was set to 0.8. The paper reports that more than 75% of cooperative mobile nodes were correctly classified across the three categories.
The noise test is less clear—and more revealing
The feedback-noise analysis needs careful reading. The document lists noise levels of 0%, 1%, 10%, 20% and 30%, then gives an accuracy sequence of 94.2%, 91.5%, 89.3% and 87.1%. Because the number of stated noise levels does not match the number of reported accuracy values, the document does not make clear which accuracy belongs to which level.
The paper nevertheless reports a clear warning at the upper end of the test: when noise exceeds 10%, malicious servers can be incorrectly identified as honest. That finding limits how confidently the reported classification performance can be extended to more heavily manipulated or unreliable feedback.
In a separate comparison, simulations that included feedback-based incentives reported a 12–18% higher cache hit ratio and throughput than scenarios without those incentives. The result is a comparison between simulated settings, not a measurement of gains for users in a live network.
The efficiency numbers do not line up neatly
The resource figures are another reason to treat the findings as preliminary. An earlier simulation summary reported CPU usage of 3–5%, memory overhead below 20 MB and average message latency of 2.1 seconds. A separate cost test, repeated 50 times with feedback data containing 10% noise, reported average CPU usage of 28.6%, memory usage of 62.4 MB and transaction latency of 0.42 seconds.
The supplied analysis notes that the paper does not reconcile those two resource summaries. They may reflect different test conditions, but the document does not provide enough detail to turn them into one consistent estimate of the system’s computational cost.
Sensitivity tests placed the best initial trust threshold at approximately 0.1, with reported detection accuracy of 94.3%, a false-positive rate of 3.7% and convergence time of 0.42 seconds. The trust-update weight was reported as stable from 0.05 to 0.2 and optimal at 0.1, with the same 94.3% detection accuracy, trust-value variance of 0.008 and convergence time of 0.42 seconds.
The pricing scenarios varied the share of low-quality edge servers across 20%, 40%, 60% and 80%. As that share rose, the authors reported an association with higher prices for high-quality servers, although the text did not provide numerical price estimates.
What the numbers do not establish
The study has no conventional control arm. Its comparisons include alternative parameter values, scenarios without feedback-based incentives and two cited schemes, including the 10% noise comparison with Xu et al. and Zhang et al. That design can show how the reported configurations differed in simulation, but it cannot isolate a causal effect of blockchain, authentication or incentives.
The reported percentages are computational outputs rather than estimates accompanied by confidence intervals or significance tests. The supplied analysis reports no formal power or sample-size rationale, and no inferential statistical model was used to quantify uncertainty around the parameter choices or performance differences.
The evaluation is entirely simulation-based, with assumptions about network composition, mobility, parameter ranges and attack conditions. The security discussion mainly describes the proposed defenses; it does not report attack-success rates from a separate adversarial experiment. Broader baseline comparisons and independent replication were also not reported.
The preprint therefore offers a model-level performance report, not evidence that the framework improves outcomes in a deployed mobile network. Its strongest reported signal is that the proposed configuration scored better than the two listed comparators under one stated noise setting, while its main unresolved issues are performance under heavier feedback noise, the inconsistent cost summaries and behavior in larger or real-world systems.
Paper data and sources
Original title: Trustworthy mobile edge caching: a blockchain approach to mitigate malicious nodes and incentivize cache sharing
Authors: Motahare Ebrahimi, Nastooh Taheri Javan, Seyedakbar Mostafavi, Fatemeh Pakzaban
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text