Internet of Things (IoT) applications are increasingly being adopted for innovative and cost-effective services. However, the IoT devices and data are susceptible to various attacks, including cyberattacks, which emphasizes the need for pervasive security measure like trust evaluation on the fly. There exist several IoT numerical data trustworthiness measures which are based on the quality of information (QoI) and correlations. The QoI measurement techniques excessively exploit heuristics, while the correlation-based approaches predict temporal correlation using an average or moving average, which limits their efficacy. To improve accuracy and reliability, we propose a model for assessing trust of IoT sensor numerical data by representing the temporal correlation using temporal relationship. We represent the temporal relationship between data within a time window in two ways: first, using the discrete cosine transform (DCT) coefficients of daily data; and second, to obtain the impact of shuttle variation, we further divide the daily data into some time windows and calculate the average of each DCT coefficient over all time windows. These two feature sets are then used to develop two independent deep neural network models. The model outcomes are fused by the Dempster-Shepard theory to calculate trust scores. The strength of our model is evaluated using both trustworthy and untrustworthy data-the former are collected from sensors under controlled supervision in a smart city project in Melbourne, Australia and the latter are generated either by simulating breached sensors or perturbing real data. Our proposed approach outperforms a contemporary correlation-based approach in terms of trust score accuracy and consistency.

中文翻译：

---

使用时间相关性的物联网传感器数值数据信任模型


物联网 (IoT) 应用越来越多地被用于创新和经济高效的服务。然而，物联网设备和数据容易受到包括网络攻击在内的各种攻击，这强调了对动态信任评估等普遍安全措施的需求。存在多种基于信息质量 (QoI) 和相关性的物联网数值数据可信度度量。 QoI 测量技术过度利用启发式方法，而基于相关性的方法则使用平均值或移动平均值来预测时间相关性，这限制了其功效。为了提高准确性和可靠性，我们提出了一种模型，通过使用时间关系表示时间相关性来评估物联网传感器数值数据的信任度。我们通过两种方式表示时间窗口内数据之间的时间关系：首先，使用每日数据的离散余弦变换（DCT）系数；其次，为了获得穿梭变化的影响，我们将每日数据进一步划分为一些时间窗口，并计算所有时间窗口上每个 DCT 系数的平均值。然后使用这两个特征集来开发两个独立的深度神经网络模型。模型结果通过 Dempster-Shepard 理论融合来计算信任分数。我们的模型的强度是使用可信和不可信的数据来评估的——前者是在澳大利亚墨尔本的一个智慧城市项目的受控监督下从传感器收集的，后者是通过模拟被破坏的传感器或扰乱真实数据来生成的。我们提出的方法在信任评分的准确性和一致性方面优于当代基于相关性的方法。