- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper focuses on a novel issue: Different clients have different probabilities to quit training, and further propose PREMIUM, a framework that prepares each client in proportion to its hazard of deletion. Experiments demonstrate the effectiveness of proposed methods.
Critical strengths
- The paper identifies that clients have different probabilities of issuing future deletion requests, while uniform preparation may waste the limited utility budget on low-risk clients. Its a novel perspective beyond conventional request-time unlearning.
- The experiments are comprehensive and convincing. The ablation studies and scalability experiments further support the proposed design.
Critical weaknesses
-
The deletion cost metrics and baseline comparisons make me confused. Eq. (4) optimizes the expected total cost over a compliance period, but the reported 110 steps are sometimes described as the cost per request. These are different metrics. Table 6 also reports 652 deletion steps when all deletion hazards are zero, which needs explanation.
-
The practical feasibility of request-time unlearning is unclear. Section 3.4 requires gradient ascent on the departing client's loss, but the server normally has no access to client data, and the client may already be offline. The authors should clarify how this operation is performed.
-
The theoretical analysis is established based on strong assumptions. Eq. (3) assumes a linear relationship between deletion cost and shared exposure, but the exposure identity in Eq. (8) does not establish this relationship for actual unlearning steps. Moreover, Assumption 2 in Appendix C.1 already assumes a similar bound, making the theoretical guarantees somewhat dependent on the assumed cost model.
Questions for authors
Please see weaknesses above.