Backdoor attacks, which manipulate deep neural networks (DNNs) by embedding specific triggers into images, have emerged as a significant security threat. Current white-box defense strategies, which are highly contingent upon access to model parameters and clean datasets, exhibit limited applicability in real-world scenarios. In contrast, black-box defense approaches demonstrate greater alignment with practical requirements. However, current black-box defense strategies primarily focus on local triggers within specific regions, which can only defend against specific types of attacks. Therefore, there is an urgent need for a black-box defense framework that can eliminate both local and global triggers while preserving semantic integrity. Recent research has elucidated that triggers in backdoor attacks predominantly manifest as high-frequency components in the frequency domain. This finding was elucidated through extensive experiments, which revealed that existing triggers are highly sensitive to image compression techniques such as JPEG compression, bit-depth reduction, and smoothing. In this paper, a black-box defense method, Frequency-Aware Purification (FAP), is proposed. FAP focuses on image purification at test time, employing techniques such as JPEG compression, Randomized Input Sampling for Explanations (RISE), and global noise addition to localize and eliminate both local and global triggers in real-time by calculating the structural similarity (SSIM) and label consistency between recovered and original samples. Additionally, FAP employs diffusion models to recover semantic information lost due to image transformations, thereby converting poisoned samples into clean ones. Extensive experimental results validate the effectiveness and generalizability of FAP across diverse backdoor attack scenarios.