Abstract: A system and method for adversarial vulnerability testing of machine learning models is proposed that receives as an input, a representation of a non-differentiable machine learning model, transforms the input model into a smoothed model and conducts an adversarial search against the smoothed model to generate an output data value representative of a potential vulnerability to adversarial examples. Variant embodiments are also proposed, directed to noise injection, hyperparameter control, and exhaustive/sampling-based searches in an effort to balance computational efficiency and accuracy in practical implementation. Flagged vulnerabilities can be used to have models re-validated, re-trained, or removed from use due to an increased cybersecurity risk profile.

Abstract: Systems and methods for diagnosing and testing fairness of machine learning models based on detecting individual violations of group definitions of fairness, via adversarial attacks that aim to perturb model inputs to generate individual violations. The systems and methods employ auxiliary machine learning models using a local surrogate for identifying group membership and assess fairness by measuring the transferability of attacks from this model. The systems and methods generate fairness indicator values indicative of discrimination risk due to the target predictions generated by the machine learning model, by comparing gradients of the machine learning model to gradients of an auxiliary machine learning model.