Uncertain cyber-attack detection and effective recovery after attacks on in-vehicle sensors are a challenge in the domain of the Internet of Vehicles (IoV). Currently, modern automatic vehicles are equipped with a variety of low-cost sensors with relatively limited computation capabilities, which constrains state-of-the-art solutions. As a result of the embedded intelligence chip in the IoV devices, deep learning algorithms can be applied to provide a novel solution for sensor attack detection and recovery. This paper is concerned with the resilient recovery after attacking the in-vehicle reference-free sensor. Toward this aim, we firstly define a theoretical model to describe automatic ground vehicles with vehicular yaw rate and local state. Then, classical attack models are established to transfer the challenge to a convex optimization issue. Finally, we consider evaluating the performance of Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and LSTM optimized by CNN (CNN-LSTM) for the general in-vehicle sensor units attack detection and recovery. The evaluation metrics for comparison experiments are accuracy, Receiver Operating Characteristic (ROC) curve, and the confusion matrix. The proposed CNN-LSTM model can obtain 95% testing accuracy, which is better than 56% for CNN and 79% for LSTM in detecting attacks. The predicting experimental results support that the LSTM is suitable for recovering signals of in-vehicle sensors.