In recommendation, the accuracy of Click-Through Rate prediction (CTR) significantly impacts user experience and economic benefits. Effective mining of hard samples is crucial for improving model performance. However, existing CTR models often perform poorly when handling hard samples due to their complexity and low-frequency feature interactions. This paper proposes a new hard sample mining method called the Diffusion-Augmented and Kolmogorov-Arnold Fourier Based Hard Sample Mining (DKAF), which integrates the advantages of the diffusion model and the Kolmogorov-Arnold Fourier Network (KAFN) to enhance the model’s ability to capture feature interactions in hard samples. The diffusion model is employed to smooth feature distributions and enhance the model’s robustness to hard sample features. A Gating Unit is used to distinguish between hard and easy samples and adopt differentiated processing strategies. KAFN replaces the traditional Multi-Layer Perceptron (MLP), as KAFN can more flexibly capture complex interactions between hard sample features through learnable activation functions. Experimental results on multiple datasets demonstrate that the DKAF model outperforms mainstream CTR models in both hard sample identification and overall performance.