SLAP — Simple Linear Attack against Perceptron (SLAP)

A novel type of attack against perceptron-based neural networks with piecewise linear activation functions using basic linear algebra is proposed. The attack is formulated as a system of linear equations and inequalities and demonstrates a simplified and computationally efficient approach to generating diverse sets of adversarial samples. The proposed attack algorithms are implemented in code available in an open-source repository. The study highlights the serious challenge that the proposed attack methodology poses to modern neural network defense emphasizing the urgent need for innovative defense strategies. By comprehensively exploring adversarial vulnerabilities, this study contributes to improving the adversarial robustness of machine learning thus paving way for developing more robust and trustworthy AI systems in real-world applications.