Customers are relying on reviews to make a purchase decision. Fake reviews are untruthful reviews written to mislead customers and can cause financial loss to Electronic Commerce (EC) businesses. Despite the importance of fake review detection, few Arabic studies exist due to the lack of suitable datasets. Therefore, we introduce the first golden-standard Arabic dataset for fake reviews detection with review content in Saudi and Modern Standard Arabic(MSA) dialects for multi-domains, restaurants, hotels, and products. We integrate the FastText word embedding technique with four Deep Neural Networks (DNN) models to study our dataset in single-domain, multi-domain, and cross-domain experiments, which showed a solid performance. Our experimental results showed CNN hybrid models are superior. The highest achieved ac- curacy was 93.71% by CNN+Bi-LSTM in the restaurant domain with a recall of 97.18%. For cross-domain, we tested the impact of zero-learning and Transfer Learning (TL) by introduc- ing ResModel, a CNN+Bi-LSTM model built on the restaurant domain dataset, the ResModel model is used further to build RHModel on the hotel domain dataset and RPModel on the prod- uct domain dataset. We showed the effect of the TL approach in improving model performance in our limited dataset, as RHModel outperformed the four DNN models in the hotel domain with an accuracy of 98.39% and a precision of 100%, we conclude that the limited size of the dataset does not result in poor performance, as the quality and relevance of dataset instances play a major role in the model performance, and to the best of our knowledge, this is the first Arabic paper that utilized a TL approach in fake reviews detection.