Abstract: Distributed acoustic sensing (DAS) using fiber-optic cables enables much denser earthquake monitoring than geophone-based systems. However, its high noise levels can obscure acquired signals, thereby affecting downstream works such as earthquake detection and phase identification, highlighting the need for noise attenuation. Classical filtering methods (e.g., bandpass and median filtering) struggle to balance noise suppression and signal preservation: conservative settings leave residual noise, while aggressive ones distort signals. Supervised deep learning requires clean labels and suffers from domain gaps between synthetic and real data. We propose a robust DAS denoising framework based on unsupervised deep learning which eliminates the need for clean reference data. A cascaded attention filtering neural network (CAFNet) is designed, consisting of two sub-networks equipped with fully-connected layers, learnable activations, and attention mechanisms. The CAFNet is optimized using a robust loss function based on a softplus-smoothed log-cosh formulation. Experiments on both synthetic and field DAS datasets demonstrate that the proposed method outperforms the benchmark methods, achieving superior noise suppression and improved signal preservation. The enhanced DAS datasets facilitate more accurate earthquake arrival picking.