- Rating
- 1
- Confidence
- 3(−1)
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper studies whether a malicious client in federated learning can undo another client's deletion request by training the unlearned model further. The authors propose Controlled Oracle Recovery, an audit with two parts. The first part infers which records were deleted from the loss change between the pre- and post-unlearning models. The second part gives the attacker the exact deleted records and checks whether training on them restores the forgotten behavior more than compute-matched training on other data. Experiments use TOFU with Llama-3.2-1B and Llama-3.1-8B and cover six unlearning methods. They show that deletion targets are often easy to identify, and that exact-target updates beat all matched controls after 100 steps, sometimes while keeping over 95% of the utility.
Critical strengths
-
The question is timely. Unlearning is usually evaluated right after the request. The paper rightly points out that in federated settings the model keeps training, so a participating client may try to undo the deletion.
-
The control design is clean. Every recovery arm starts from the same unlearned state with the same optimizer and number of steps, and only the training data changes. This is a sensible way to separate target-specific recovery from ordinary training.
-
The evaluation covers six unlearning methods, two model sizes and two forget ratios. It reports recovery together with retained utility, and the recovery-utility frontiers in Appendix E are a useful view.
Critical weaknesses
-
The most important point, and the one that confuses me most, is that there are no federated experiments. Despite the title and Figure 1, all experiments are run in a centralized setting, with no aggregation, no other clients and no later rounds. The conclusion itself admits that the effect after aggregation remains to be tested. As a result, the main claim about federated unlearning is not supported by the experiments.
-
The main finding is largely predictable. Fine-tuning on the exact forget records directly optimizes the forget-answer probability, which is the very metric being measured. So it is not surprising that this beats training on retain, holdout or fluent text. Prior relearning work [1] has already shown recovery with much weaker access.
-
The key control is missing. A model that never saw the forget data could also learn it from the exact records. In Figures 3 to 6, the gold-retrain curves appear to reach a similar level after 100 steps for several methods. Without the retrained model as a control, the results can't tell residual knowledge apart from simply relearning the data.
-
The threat model is toostrong. The recovery arm receives the exact deletion set as an oracle, separately from the inference step. An attacker who already holds these records could teach them to any model, even a retrained one.
-
The deletion-target inference is not new, and it doesn't indicate an unlearning failure. Using the pre/post loss change follows Chen et al. (2021). Gold retraining also reaches a ROC-AUC of about 0.999, so high separability is expected for any method that changes the model on the forget set, including exact retraining.
-
The evaluation is narrow and the statistics are incomplete. Recovery is only tested on TOFU, with a single learning rate and a single unlearned state per method.
-
Parts of the paper look unfinished? There is a broken reference, and labels such as "Eq. equation 9" appear repeatedly. The appendix figures also use very small fonts and inconsistent styles.
Questions for authors
-
Could the authors run the attack in an actual federated setting, with aggregation, update clipping and several rounds? Does the recovery survive aggregation?
-
Could the authors apply the same exact-data updates to the gold-retrained model and use it as a control?
-
What happens when recovery uses the inferred candidates from RQ1 instead of the oracle set, or uses only paraphrased or partial forget data?
-
How sensitive are the results to the client numbers, learning rate and the number of steps?
-
How does this work differ from Sheng's recent work [2], which also studies retaliatory attacks against federated unlearning?
-
Could the authors test recovery on another benchmark, such as MUSE or WMDP?
[1] Hu, Shengyuan, et al. "Unlearning or obfuscating? jogging the memory of unlearned llms via benign relearning." International Conference on Learning Representations. Vol. 2025. 2025.
[2] Sheng, Xinyi, et al. "Retaliatory attacks against federated unlearning via data leakage." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 30. 2026.
Thank you for the response and clarifications. I appreciate the additional WMDP experiments and the planned learning-rate analysis. However, I still have a few remaining concerns:
First, coz of lack of aggregation or subsequent federated rounds, the experiment in the paper could only demonstrate local recovery, rather than a successful attack on the global federated model.
Meanwhile, since the evaluation does not distinguish between seen and unseen forget samples, it is difficult to determine whether the observed recovery is from the residual knowledge or relearning from the provided examples.
And for gold-retrained control, thanks for pointing out that this control is included in appx. However, the current comparison seems to have limited statistical support.
Additionally, I did not see the revisions in the updated paper or a more comprehensive discussion addressing these concerns. Based on what mentioned above, I have to keep my original score but lower my confidence to 3. I would also be willing to reconsider my assessment if further experiments could address the issues.