Abstract:

Machine learning typically mines underlying patterns and rules from data, making it susceptible to phenomena such as overfitting and underfitting, which in turn affects the generalization and robustness of learning models. This paper explores the potential fragility and instability of SVM from the perspective of adversarial simulation testing. The adversarial simulation strategy employed involves selectively contaminating training sample labels to simulate an attack on the SVM classifier, thereby degrading its performance and testing its dependency on training samples. To explore the ceiling of performance degradation of an SVM classifier under the combination attack of different samples, the contradictory objectives of minimum attack cost-maximum attack effectiveness are designed, and a multi-objective optimization model is constructed for SVM simulation tests. This model is fundamentally a typical multi-objective combinatorial optimization problem that can be properly solved using multi-objective evolutionary algorithms to find a set of non-dominated solutions among the objectives, facilitating the investigation of the classifier's stability under the combination attack of different samples. Comparative experimental results on simulated and real datasets show that the proposed method can identify optimal attack sample combination at varying attack levels in a single run, and achieve more severe classification performance degradation, making it more suitable and effective for investigating the stability of classifiers comprehensively.

Key words: adversarial simulation testing, label contamination, SVM, performance degradation, multi-objective optimization, non-dominated solutions