Modern Global Navigation Satellite System (GNSS) architectures incorporate authorized signals like GPS M-code; however, conventional extraction methods suffer from performance degradation under low signal-to-noise-ratio (SNR) conditions and exhibit strong dependence on high-gain antennas and precise synchronization. This paper proposes a machine learning-based end-to-end extraction framework leveraging time-frequency domain feature fusion. This method breaks through the constraint of relying solely on either time-domain or frequency-domain features. It jointly feeds the time-domain waveforms and spectral features of baseband signals into models such as Multi-Layer Perceptron (MLP) and Transformer, enabling automatic learning of the nonlinear time-frequency characteristics of M-code. This approach effectively suppresses interference from P(Y) code sidelobes and fully exploits the information contained in both the main and side lobes of the M code spectrum. Experimental results demonstrate that under the extremely low SNR condition of −10 dB, the extraction accuracy of the proposed method is improved by 13.7% compared with conventional methods. Systematic accuracy–efficiency trade-off analysis shows that the lightweight MLP model achieves comparable accuracy to the complex Transformer model, with only 6.8% of the parameter scale and 6.2 times faster inference speed, making it the most competitive solution for real-time engineering deployment. In the real-world measurement scenario using a 7.5-m antenna, an extraction accuracy of 94.7% is achieved with only 10 ms of small-sample training data. This method significantly enhances the extraction performance of authorized signals under low-SNR non-cooperative reception conditions.