<p>Secure Code Estimation and Replay (SCER) spoofing attacks manipulates Global Navigation Satellite System (GNSS) signals by exploiting estimated encrypted PRN codes to spoof non-cooperative receivers, serving as a critical enabler for countering rogue or unauthorized unmanned aerial vehicle (UAV) operations. Therefore, precise non-cooperative code estimation directly influences the effectiveness of SCER spoofing attacks. However, existing code estimation methods face the challenge of limited estimation accuracy due to signal-to-noise ratio (SNR) constraints. The non-uniform constellation phases of modern GNSS complex modulation also make code estimation even more difficult. To address these issues, this paper proposes a deep learning (DL) approach named TurboUNet. Turbo codes leverage the probabilistic outputs of each iteration as prior information for the subsequent estimation, effectively enhancing the information exchange and decoding capabilities across the modules. The UNet employs gated units to perform weighted fusion of detail-rich low-level features extracted by the GNSS signal and progressively refines the code estimates throughout the decoding process. By combining the iterative decoding characteristics of Turbo codes with the refined encoding–decoding structure of UNet, the proposed method incrementally improves the code estimation accuracy, overcoming the limitations imposed by low SNR and non-uniform constellation. Simulation and practical experiments demonstrate that the proposed method consistently outperforms conventional approaches in terms of BER under varying SNRs and sampling rates. In practical measurements, the proposed method achieves up to 22% and 27% BER reduction for P(Y) and M codes, respectively, over conventional approaches, and up to 5.5% and 6% over comparative DL methods.</p>