The Price of Ignorance: Information-Free Quotation for Data Retention in Machine Unlearning

A team of researchers including Bin Han, Di Feng, Zexin Fang, Jie Wang, and Hans D. Schotten has published a paper titled 'The Price of Ignorance: Information-Free Quotation for Data Retention in Machine Unlearning' on arXiv. The work addresses a critical conflict in data regulation: when users exercise GDPR 'right to be forgotten' requests, mobile network operators face a dilemma. Excessive machine unlearning (the process of removing specific data points from trained AI models) degrades model accuracy and incurs significant retraining costs. However, existing pricing mechanisms for data retention require servers to know each user's private privacy and accuracy preferences—information that's fundamentally unavailable under the very regulations that mandate unlearning.

The researchers' breakthrough is an 'information-free ascending quotation mechanism' where the server simply broadcasts progressively higher prices, and users self-select their data supply. This elegant solution requires no knowledge of users' private parameters. The team formalizes the 'Price of Ignorance'—the welfare gap between optimal personalized pricing (which knows everything) and their information-free approach (which knows nothing)—and proves a three-regime efficiency ordering. Their numerical evaluation across seven different mechanisms and 5000 Monte Carlo simulation runs shows this price is remarkably low: the information-free mechanism achieves 99% or more of the welfare of information-intensive benchmarks.

Beyond near-perfect efficiency, the proposed system provides noise-robust guarantees and comparable fairness metrics to traditional approaches. This represents a significant advancement for practical implementation of machine unlearning in regulated environments, offering a viable path for companies to comply with data deletion mandates while maintaining model performance and economic viability. The mechanism's simplicity makes it particularly suitable for large-scale deployment where collecting individual preference data would be both impractical and legally problematic.