In recent years, deep learning based malicious traffic detection (MTD) systems have demonstrated remarkable success. However, their effectiveness tend to decrease because most malicious traffic datasets are suffered from noisy-labeled and long-tailed problems. While numerous approaches have been developed to address these two problems individually, they become inefficient when confronting the combined challenge of noisy long-tailed data, as they typically tackle only a single adverse factor at a time. This paper proposes a two-stage method called Distribution-aware sample Selection and Dynamic Instance-based Relabeling (DSDIR), which simultaneously addresses the impacts of noisy-labeled and long-tailed problems. In the first stage, a noise-independent clean sample selection method is designed to obtain a clean dataset, which converts negative effects of the long-tailed problem into positive ones. In the second stage, dynamic instance-based relabeling is designed to train a model and improve the dataset’s quality simultaneously. Eventually, DSDIR not only produces a balanced and noise-tolerant model but also obtains a clean dataset. Experimental results demonstrate that in the high noise condition of 60% and 80%, the accuracy rate of DSDIR is 5% higher than the state-of-the-art methods. Our code is available at https://github.com/nku-ligl/DSDIR.

Federated learning, as a distributed machine learning paradigm, enhances privacy protection but faces the challenge of heterogeneity. Data-free knowledge distillation (DFKD) methods attempt to overcome this challenge by using a generator to synthesize samples for fine-tuning a global model. However, these methods often suffer from significant shifts in output distribution, leading to catastrophic forgetting. To tackle these issues, we propose FedFLD, a novel federated forget-less distillation framework that mitigates catastrophic forgetting in DFKD while addressing the heterogeneity challenge. Specifically, FedFLD guides the generator’s training from three key aspects and constrains the output distribution with an elastic weight consolidation penalty term. By synthesizing diverse samples from different perspectives through additional generator updates, FedFLD facilitates effective knowledge distillation from local models to the global model. Additionally, the global model is further optimized via the heterogeneity fine-tuning process, mitigating the bias from heterogeneity and resulting in a more expressive and robust model.