In the black-box attack scenario, traditional methods typically adopt either spatial domain transformations or frequency domain perturbations independently to enhance model perturbation strategies, but seldom explore the complementarity between gradient directions in the frequency and spatial domains. To address this, this paper proposes an Orthogonal Frequency-Spatial Gradient Fusion Attack for Boosting Adversarial Transferability (OFSGF), which decomposes the spatial domain gradient ( \(g_s\) ) into parallel components ( \(g_{s\parallel }\) )) and orthogonal components ( \(g_{s\perp }\) ) in the direction of the frequency domain gradient ( \(g_f\) ), and dynamic fuses ( \(g_f\) ) and ( \(g_{s\perp }\) ) to cover more comprehensive disturbance directions. The orthogonal component is used to supplement the model sensitive features that are not sufficiently optimized by frequency domain perturbations. We design a three stage strategy as follows: first, the Channel Adaptive Enhancement (CAE) method is employed to perform differentiated modulation of color channels in the spatial domain and Combination perturbations with adaptive Gaussian noise; second, the Dynamic Frequency Mask (DFM) is used to identify key frequency components, generate adaptive masks, and introduce random perturbations; finally, we propose a frequency spatial gradient orthogonal fusion mechanism, which decomposes the spatial gradient into two components parallel and orthogonal to the frequency domain gradient and fuses the orthogonal component with a dynamic weight \(\alpha \) , introducing cross-model generalization capability while preserving the dominant direction in the frequency domain, thus alleviating the problem of gradient direction overfitting. Experiments show that, compared with baseline methods, our method improves the average attack success rate by 8.9% on both standard and defense models.