Poisoning Attack Against Estimating From Pairwise Comparisons.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Pairwise ranking algorithms are widely used in various applications like elections, sports, and information retrieval.
• Data poisoning attacks, where malicious data is injected to manipulate models, pose a significant threat to these algorithms.

Purpose of the Study:

• To systematically investigate data poisoning attacks specifically targeting pairwise ranking algorithms.
• To develop computationally tractable methods for analyzing and executing such attacks.

Main Methods:

• Formalized data poisoning attacks as dynamic and static games, modeled as integer programming problems.
• Reformulated integer programming problems into distributionally robust optimization (DRO) problems for tractability.
• Proposed two efficient poisoning attack algorithms based on DRO formulations.

Main Results:

• Established theoretical guarantees for the proposed attack algorithms, including Nash equilibrium existence and generalization bounds.
• Demonstrated significant performance degradation of ranking algorithms through simulations and real-world experiments.
• Showcased a dramatic decrease in the correlation between true and attacked ranking lists.

Conclusions:

• The study provides the first systematic analysis of data poisoning attacks on pairwise ranking.
• The developed DRO-based methods offer efficient and theoretically grounded strategies for launching effective attacks.
• These findings highlight the vulnerability of pairwise ranking systems and the need for robust defense mechanisms.