Adaptive Adversarial Training Does Not Increase Recourse Costs

Adaptive adversarial training may offer a solution to a key trade-off in responsible AI: Traditional adversarial training (which makes models resistant to malicious attacks) has been shown to increase the cost of algorithmic recourse—the minimal changes needed for individuals to overturn unfavorable AI decisions. This paper finds that adaptive adversarial training, which adjusts robustness requirements based on individual instances, does not significantly increase recourse costs, suggesting a path forward for systems that are both robust and fair.

The findings challenge previous assumptions about the robustness-recourse trade-off: Prior research established that larger adversarial training radii correlate with higher recourse costs, creating a dilemma for practitioners. This work demonstrates that adaptive approaches can improve model robustness "with little effect on algorithmic recourse costs," potentially resolving this tension in domains where both security and individual recourse are important.

The research addresses a practical gap in adversarial training implementation: The paper notes that "the appropriate adversarial training radius has always been unknown," highlighting a real implementation challenge. By showing that adaptive methods work well without specifying a fixed radius, this research provides practical guidance for deploying more robust AI systems in real-world applications.

This work is particularly relevant for high-stakes decision systems: The focus on domains "where recoursability is critical" suggests applications like lending, hiring, or criminal justice—contexts where individuals need meaningful ability to challenge and correct unfavorable algorithmic decisions while the system remains secure against manipulation.