Data fusion is the critical process for data transmission in the Internet of things (IoT). In data fusion process, the original sensing data is processed and aggregated in the network, and only the aggregated results of data fusion are sent to the application layer, which effectively reduces the resource consumption and alleviates the workload on the sink node. Since no network node caches the aggregated data in the data fusion process, it is impossible to detect and locate the false data injection attack against data fusion results. In order to mitigate this significant vulnerability, an efficient detection scheme of false data fusion is proposed in this paper. By modeling the data fusion process, we discover and model the relationship between the input data and the fusion results, and apply the obtained model to detect abnormal data fusion results. In this way, we can mitigate malicious data fusion and optimize the IoT transmission security. In detail, we first collect input data and the relevant data fusion results for each node, and a compressed feature representation mechanism is designed to improve the data collection efficiency and reduce the resource consumption. In addition, a data fusion model based on probabilistic graph model is proposed to depict the spatial and temporal relationship between the input data and the data fusion results. Ultimately, we take the model to detect the abnormal data fusion results in an efficient way. The experimental results demonstrate that the proposed detection scheme can detect malicious data fusion operations efficiently and accurately and thus guarantee IoT transmission security.