Multi-view self-supervised learning for label-flipping robustness in tabular data: a comparative study

Tabular machine learning models are widely used in high-stakes domains such as finance, healthcare, and network intrusion detection, where their reliability depends heavily on accurate training labels. However, malicious or accidental label flipping can corrupt training data and degrade downstream model performance. Recent attacks further target decision-boundary or vulnerable local regions rather than flipping labels randomly. This paper presents a BYOL-based multi-view KNN sanitization approach for identifying suspicious label-flipped samples in tabular data. The method learns label-independent representations and examines KNN label-disagreement patterns across original and perturbed SSL embedding views. We evaluate the method on six cross-domain tabular benchmarks under both large-scale and small-scale label-flipping attacks, and compare it with existing sanitization baselines and a SCARF-based representation baseline. The results show that multi-view neighborhood probing improves suspicious-label detection and provides a practical F1-oriented trade-off, while downstream recovery remains attack- and target-model-dependent. These findings provide actionable insights for improving the robustness and label integrity of tabular machine learning systems against label-flipping attacks.