Efficient Detection of Black Box Adversarial Attacks in Machine Learning Systems Using Cluster-Based and Class-Based Input Patterns

Adversarial attacks pose a significant challenge to the security and reliability of machine learning models. This paper focuses on detecting black-box adversarial attacks, where attackers operate without access to the model's internal parameters or architecture—a scenario more relevant to real-world threats than white-box settings, where such knowledge is assumed. Two novel detection methodologies are introduced: (1) a clustering-based approach that groups input data to identify anomalies and (2) a class-centric method that uses statistical properties of labeled data to compute class centers and boundaries for detecting deviations. These methods utilize thresholds derived from dynamic statistical metrics to ensure robust, efficient, and scalable detection.Extensive experiments validate the effectiveness of these approaches using the Iris dataset. Results demonstrate that the clustering-based method excels in detecting adversarial samples, achieving high accuracy, while the class-based method offers a conservative alternative by minimizing false positives. Both methodologies provide practical solutions for enhancing model security, particularly in scenarios where model details are unavailable, thereby addressing the limitations of existing defense mechanisms. Future work will explore scalability and application to higher-dimensional datasets.