Fig.1.

Overview of enhancement methods used in this paper. Classification accuracy in the original dataset is 50%. After applying gradient-based enhancement attacks following Algorithm 1, classification accuracy in the enhanced dataset is 99%. Using method enhancement attacks (Algorithm 2), datasets are altered such that a specific method (e.g., feedforward neural network) outperforms another (e.g., support vector machine). In all cases, the changes between the original and enhanced datasets are minor. The “Downstream effects” box highlights possible implications of enhancement attacks.