Adversarial vulnerabilities of human decision-making.

Area of Science:

• Cognitive Science
• Neuroscience
• Artificial Intelligence

Background:

• Adversarial examples are inputs designed to fool artificial and natural neural networks.
• Human learning and choice processes may exhibit similar vulnerabilities to adversarial manipulation.

Purpose of the Study:

• To examine adversarial vulnerabilities in human learning and choice.
• To propose a framework for generating adversarial opponents to shape human choices.
• To investigate the strategies adversaries use to exploit human decision-making vulnerabilities.

Main Methods:

• Utilized recurrent neural network models of choice processes.
• Developed a general framework for creating adversarial opponents.
• Conducted three experiments on action selection, response inhibition, and social decision-making.

Main Results:

• Demonstrated the efficacy of the adversarial framework in influencing human choices across different tasks.
• Identified specific strategies employed by adversaries to exploit human decision-making.
• Provided empirical evidence for adversarial vulnerabilities in human cognitive processes.

Conclusions:

• Human learning and choice are susceptible to adversarial manipulation.
• The proposed framework offers a method to study and potentially mitigate flawed decision-making.
• Findings have implications for understanding and intervening in behavioral patterns across the behavioral sciences.