Abstract
The label flipping attack is a special poisoning attack in the adversarial environment. By adding noise to the data, it destroys the learning process of the model and affects the decision-making performance of the model. Recent literature work uses semi-supervised learning techniques to defend against label flipping attacks. However, these methods require a clean dataset to achieve their goals. This study proposes a novel label noise processing framework to correct the labels of contaminated samples in the data preprocessing stage. Based on five real UCI datasets, we evaluated the effectiveness of the semi-supervised defensive label noise correction algorithm based on the AdaBoost algorithm (AdaSSL). With a noise ratio of 0~20%, we compared the classification performance of six classic machine learning algorithms (NB, LR, SVM, DT, KNN and MLP) under the AdaSSL defense algorithm. The results show that compared with the most advanced semi-supervised defense algorithms in the literature, our algorithm does not need to use an additional dataset. At a noise ratio of 10%, the AdaSSL algorithm is significantly better than state of the art label noise defense technology.
Introduction
In machine learning, the integrity of training data is paramount. Label flipping attacks represent a sophisticated form of adversarial poisoning where an attacker intentionally alters training labels (e.g., mislabeling spam as legitimate). Such attacks corrupt the learning process, leading to degraded model performance and unreliable decision-making. While existing defenses often rely on a "gold standard" clean dataset for guidance, obtaining such data is frequently impractical. This paper introduces AdaSSL, a framework designed to detect and correct label noise using only the contaminated data available.
Methodology: The AdaSSL Framework
The proposed AdaSSL (AdaBoost-based Semi-Supervised Learning) framework operates during the data preprocessing stage through a two-step process:
- Noise Identification via AdaBoost: AdaBoost iteratively trains weak classifiers, assigning higher weights to samples that are difficult to classify. In a poisoned dataset, these high-weight samples typically correspond to flipped labels where features and labels are inconsistent. The framework partitions the data into a "Weight_Max Set" (likely noisy) and a "Weight_Min Set" (likely clean).
- Correction via Semi-Supervised Learning: Rather than discarding suspicious samples, AdaSSL treats the "Weight_Max Set" as unlabeled data. It utilizes the "Weight_Min Set" as a reliable foundation to predict correct labels for these samples, generating a "Cleaned Set."
Experimental Setup
The effectiveness of AdaSSL was validated using five datasets from the UCI Machine Learning Repository: Spambase, Kr-vs-kp, Biodeg, Diabetes, and Breast-w. These datasets cover diverse domains, ensuring the framework's generalizability. Performance was measured across six standard machine learning algorithms: Naive Bayes (NB), Logistic Regression (LR), Support Vector Machines (SVM), Decision Trees (DT), K-Nearest Neighbors (KNN), and Multi-Layer Perceptrons (MLP).
Key Results
The experimental results demonstrate that AdaSSL provides robust protection against label flipping attacks.
- Superior Performance: At a 10% noise ratio, AdaSSL significantly outperformed existing state-of-the-art defense technologies across accuracy, F1-score, and AUC metrics.
- Stability and Consistency: As noise levels increased from 0% to 20%, models protected by AdaSSL exhibited significantly slower performance degradation compared to baseline "entropy_attack" defenses. This consistency was observed across all five UCI datasets.
Conclusion
AdaSSL offers a practical, self-contained solution for defending against label flipping attacks without requiring external clean datasets. By repurposing AdaBoost's weighting mechanism for diagnostic purposes and employing semi-supervised correction, the framework preserves training data volume while enhancing model robustness. Future work will address computational overhead for large-scale datasets and performance under extreme noise conditions (>40%).
Summary updated