Law and Adversarial Machine Learning

TL;DR

This paper examines how existing legal frameworks might respond to adversarial attacks on machine learning systems, emphasizing the need for transparency, forensic-ready design, and civil liberties considerations.

Contribution

It analyzes legal implications of adversarial ML attacks and advocates for transparent benchmarks, forensic-aware architecture, and broader civil liberties considerations.

Findings

Certain attacks are more likely to lead to legal liability.

Legal responses depend on attack types and context.

Calls for ML community to develop standardized benchmarks.

Abstract

When machine learning systems fail because of adversarial manipulation, how should society expect the law to respond? Through scenarios grounded in adversarial ML literature, we explore how some aspects of computer crime, copyright, and tort law interface with perturbation, poisoning, model stealing and model inversion attacks to show how some attacks are more likely to result in liability than others. We end with a call for action to ML researchers to invest in transparent benchmarks of attacks and defenses; architect ML systems with forensics in mind and finally, think more about adversarial machine learning in the context of civil liberties. The paper is targeted towards ML researchers who have no legal background.