Knowledge Tracing (KT) is a fundamental task in personalized online education, aiming to model students’ evolving knowledge states and predict their future performance. However, most existing KT models fail to sufficiently account for the impact of Noisy Implicit Feedback (NIF), which is prevalent in student interaction records. NIF refers to interactions that deviate from a student’s true knowledge state due to various uncertainties, including random responses, system-induced errors, guessing, and slips, thereby introducing biases into KT models and degrading their predictive performance. To tackle this challenge, we propose a general Dual-Level Denoising KT framework (DLDKT), which identifies and mitigates the impact of NIF by operating at both the interaction and knowledge state levels. Specifically, we introduce a diffusion model that leverages student embedding to design a reverse process for reconstructing personalized student interactions, facilitating the identification of noisy data within interactions. Furthermore, we model the Accumulated Knowledge State and then apply an attention mechanism to capture the locations of noisy data within knowledge states. Finally, we design an Indirect Denoising Strategy to alleviate the biases induced by NIF. Extensive experiments on real-world benchmark datasets demonstrate that DLDKT effectively mitigates the impact of NIF on prevailing KT models and significantly improves predictive accuracy.