This paper presents a deep learning-based indoor fingerprinting system called DeepFi that uses Channel State Information (CSI) for indoor localization. DeepFi is designed to improve the accuracy and efficiency of indoor positioning by leveraging the rich information provided by CSI, which includes amplitude and phase data from multiple subcarriers across three antennas. The system consists of an offline training phase and an online localization phase. During the offline phase, deep learning is used to train a deep network to extract features from CSI data, which are then used as fingerprints for localization. A greedy learning algorithm is employed to train the network layer by layer, reducing computational complexity. In the online phase, a probabilistic method based on radial basis functions is used to estimate the location of a mobile device. The DeepFi system is validated through experiments in two indoor environments: a living room and a computer laboratory. The results show that DeepFi outperforms existing methods such as FIFS and Horus in terms of localization accuracy. The system's performance is also evaluated under different conditions, including varying numbers of antennas, test packets, and packets per batch. The experiments demonstrate that DeepFi achieves better accuracy and efficiency compared to traditional methods, particularly in complex propagation environments. The system's ability to exploit the fine-grained properties of CSI data from multiple subcarriers and antennas contributes to its superior performance. The results confirm that DeepFi can effectively reduce location errors and provide accurate indoor localization in various scenarios.This paper presents a deep learning-based indoor fingerprinting system called DeepFi that uses Channel State Information (CSI) for indoor localization. DeepFi is designed to improve the accuracy and efficiency of indoor positioning by leveraging the rich information provided by CSI, which includes amplitude and phase data from multiple subcarriers across three antennas. The system consists of an offline training phase and an online localization phase. During the offline phase, deep learning is used to train a deep network to extract features from CSI data, which are then used as fingerprints for localization. A greedy learning algorithm is employed to train the network layer by layer, reducing computational complexity. In the online phase, a probabilistic method based on radial basis functions is used to estimate the location of a mobile device. The DeepFi system is validated through experiments in two indoor environments: a living room and a computer laboratory. The results show that DeepFi outperforms existing methods such as FIFS and Horus in terms of localization accuracy. The system's performance is also evaluated under different conditions, including varying numbers of antennas, test packets, and packets per batch. The experiments demonstrate that DeepFi achieves better accuracy and efficiency compared to traditional methods, particularly in complex propagation environments. The system's ability to exploit the fine-grained properties of CSI data from multiple subcarriers and antennas contributes to its superior performance. The results confirm that DeepFi can effectively reduce location errors and provide accurate indoor localization in various scenarios.