Mirage Framework Exposes Flaws in Machine Unlearning Certification
The Mirage Framework has unveiled significant vulnerabilities in the certification process for machine unlearning within vertical federated learning (VFL). This auditing framework employs four diagnostics: LPR, CKA, Feature Separability Scoring, and Layer-Wise Recovery Analysis. Experiments conducted on seven datasets and baseline methods revealed that LPR outperformed the retrained baseline by as much as 15.4 points. Additionally, CKA demonstrated that models remained more structurally similar to the original than to the retrained versions. Results indicated ongoing geometric discrimination, challenging existing output-level certification claims. The findings are detailed in a study available on arXiv, under identifier 2605.20282.
Key facts
- Mirage is a representation-level auditing framework for machine unlearning in VFL.
- It includes four diagnostics: LPR, CKA, Feature Separability Scoring, and Layer-Wise Recovery Analysis.
- Experiments were conducted across seven datasets and seven baseline methods.
- LPR exceeded the retrained baseline by up to 15.4 points.
- CKA showed models remain structurally closer to the original than to the retrained reference.
- Separability scores indicated persistent geometric discrimination.
- The study is published on arXiv with identifier 2605.20282.
- The work challenges output-level certification claims in VFL unlearning.
Entities
Institutions
- arXiv