Exploring Fairness in Educational Data Mining in the Context of the Right to be Forgotten

Machine unlearning requests can be weaponized to undermine fairness: The paper demonstrates that bad actors can submit seemingly legitimate "right to be forgotten" requests that are actually designed to make AI systems discriminate against specific groups while maintaining overall accuracy—making the bias invisible to system monitors.

Educational AI systems are vulnerable to fairness attacks: The research shows these attacks are particularly effective in educational data mining contexts, where student records and algorithmic decisions about learning pathways could be manipulated to disadvantage certain student populations without triggering performance alarms.

Current safeguards assume good faith: Existing approaches to machine unlearning operate under the assumption that data removal requests are honest and made in secure environments. This paper reveals that assumption is dangerous, as malicious actors can exploit the unlearning process itself as an attack vector.

Fairness and accuracy metrics can mask discrimination: A critical finding is that models targeted by these attacks maintain their predictive accuracy while becoming unfair—meaning organizations relying solely on accuracy metrics to validate their systems would fail to detect that bias has been introduced.

New protections are needed for the right to be forgotten: While the right to be forgotten is an important privacy protection, implementing it safely in educational AI requires new security measures to verify that unlearning requests are legitimate and don't introduce hidden discrimination.