Keywords: adversarial examples, natural langauge, large-scale surveys
Securing systems that collect and process large-scale text from online written surveys from fraudulent submissions is a critical capability, particularly for large organizations that conduct frequent, complex surveys across a large, often distributed workforce. Penetration testing of potential security flaws often requires labor-intensive manual generation of attack vectors to test the resilience of fraud detection system. Here we discuss algorithmic methods for generating sufficiently robust natural language adversarial examples to test fraud detection systems that require the submission of different categories of structured and unstructured lanauge for a stylized survey, and discuss practical methods of fusing side information, such as that obtained from traditional social engineering methods, with these attack vectors.