Exécution de notre outil de synthèse sur un nouvel article

Résumé : Synthetic data generation has emerged as a promising approach to address challenges related to data scarcity, privacy constraints, and bias mitigation in machine learning applications. In networking, especially intrusion detection systems (IDS), access to balanced and unbiased datasets is critical for developing robust models. However, real-world network traffic datasets often exhibit significant class imbalance and encode discriminatory patterns across sensitive attributes, which can result in unfair decision-making. This paper proposes FairTabFormer, a transformer-based generative model tailored to produce realistic, fair, tabular synthetic network traffic data. Based on the GPT-2 decoder architecture, FairTabFormer integrates a fairness-aware training paradigm that penalizes disparity in prediction distributions across sensitive groups, specifically targeting protocol types in intrusion detection datasets. The model leverages a custom loss function that combines cross-entropy with fairness regularization, ensuring demographic parity in the generated data. The experimental evaluation uses the CIC IoT23 dataset, a comprehensive benchmark for IoT-based threat detection. The downstream performance is assessed using an XGBoost classifier using accuracy as the primary utility metric, while fairness is measured using the Demographic Parity Ratio (DPR) and the Demographic Parity Difference (DPD). Comparative results against state-of-the-art models such as CTGAN and REaLTabFormer demonstrate that FairTabFormer achieves superior fairness while maintaining high predictive utility. These findings underscore its potential to enhance fairness without compromising data quality.