- Rating
- 4(+2)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper links signals obtained from the learning dynamics to privacy (specifically membership inference) and robustness concerns. As these signals also relate to memorizations, the paper shows how memorizations, privacy and robustness are all related to each other. The paper provides experiments to support its theory.
Critical strengths
-
The paper finds a new, cheaper method to evaluate how much a sample is vulnerable to privacy and robustness attacks. The method relies on signals obtained from the learning dynamic of the learning algorithm, so it is cheap to obtain them during training.
-
As the same signals are used to both privacy and robustness, the paper links the two topics using those signals and the notation of memorization.
-
The paper provides various experiments to support its theory and showing how to use the signals to define a metric that measures how much a sample is vulnerable to privacy attacks.
-
I really like that every theorem has an interpretation.
Critical weaknesses
-
Regarding background and assumptions - they seem really strong. Are there any models that satisfy the assumptions needed for the theorems? especially the L-bounded loss (equation 5).
-
line 234 - and are unclear to me. In the end, only the last checkpoint of the model is published, so why would one care about all the during the training phase? It seems to me that only is relevant.
-
equation 13, why this is the interesting direction of the inequalities? It seems to me that it's more interesting to prove an upper bound to and a lower bound to .
-
I don't agree with Theorem's 2 interpretation. is the weakest point during training phase, but the deployed model will have , so the upper bound may not apply to . Similary, the lower bound may not apply to since .
-
Is there a way for the attacker to estimate CSL and CSG? if not, the bounds are not exposed to the attacker and he has no additional information to exploit.
Overall, I liked the paper and I will gladly raise my score based on the rebuttal.
Questions for authors
-
Equation 7 - there is no in the equation, but rather . I guess it's a typo.
-
Section 5 - is the perturbation defined per sample, or is it global? If it's related to a specific sample , consider adding a symbol so it would be more understandable, for example .
-
Section 5 - I don't fully understand the terms "necessary threshold" and "sufficient threshold". Consider adding a short explanation to the terms. I see that they are defined in the appendix, so maybe consider adding a reference.
I want to stress that rejecting my suggestions here will not affect my final score, they are just friendly suggestions that you may or may not agree with.
4 comments
5. Is there a way for the attacker to estimate CSL and CSG? If not, the bounds are not exposed to the attacker and he has no additional information to exploit.
A: No. An external attacker cannot compute CSL or CSG, since both require the training trajectory. With white-box access to the final checkpoint, the attacker can only compute the loss and input gradient that enter the single-checkpoint bounds. This works in the defender's favour: our framework is designed as a zero-cost auditing tool for the model developer, not as an attack tool. The developer obtains CSL and CSG for free during training, can identify records vulnerable to external C&W or shadow-model MIA attacks, and can mitigate them before publishing the final checkpoint. Note that the attacker's inability to compute the bounds does not remove the vulnerability itself, since attacks such as C&W and LiRA exploit it directly, but with a significantly higher computation budget.
6. In Equation 7, there is no in the equation, but rather , which appears to be a typo.
A: Thank you for catching this. Since denotes the input-label tuple, we will explicitly write in Equation 7 to avoid confusion.
7. In Section 5, is the perturbation defined per sample, or is it global? If it's related to a specific sample , consider adding a symbol so it would be more understandable, for example .
A: Yes, , , and in Section 5 are defined per sample . We will update the notation to , , and .
8. In Section 5, I don't fully understand the terms "necessary threshold" and "sufficient threshold", so consider adding a short explanation or a reference to the appendix.
A: These are derived in Lemma 1 (App. B.2.1), and we will add the following short explanation. In -class classification with Cross-Entropy loss, loss values fall into three regimes. Below the necessary threshold (), the sample is always correctly classified because the true class holds over half the probability, so crossing is necessary for misclassification. In between (), misclassification is not guaranteed and depends on how the remaining probability is split among the other classes. Above the sufficient threshold (), the true-class probability drops below , so by the pigeonhole principle another class wins and crossing is sufficient for misclassification.
I thank the authors for their answer. All my concerns are resolved. I have raised my score to 4
We sincerely thank the reviewer for their time and efforts to read and respond to our Rebuttal.
We thank the reviewer for their time and efforts in reviewing our work. We answer the questions below.
1. The background and assumptions seem really strong; are there any models that satisfy the assumptions needed for the theorems, especially the L-bounded loss (equation 5)?
A: Theorems 1 and 2 do not require L-bounded loss, stability, or generalization. They require only smoothness and a non-negative loss l≥0, which unbounded Cross-Entropy satisfies. Corollary 1(i) additionally requires error stability β, through the memorization bound of Ravikumar et al. (2025), but not L-bounded loss. L-bounded loss is required only in Theorem 3 and Corollary 1(ii), to normalize the loss following Theorem 2 of Yeom et al. (2018).
Networks with smooth activations (e.g. softplus, GELU, SiLU) combined with softmax Cross-Entropy satisfy the smoothness assumptions for finite weights. Cross-Entropy is bounded whenever predicted probabilities stay bounded away from zero, which holds for finite weights on the bounded input domain [0,1]d. In practice, networks trained with weight decay plateau at a finite maximum loss. Nearly all the assumptions, including the L-bounded loss, were adapted from the work on CSL, CSG, privacy and memorization by Ravikumar et al. (2024; 2025; 2026), where they are justified in more detail. ReLU networks do not satisfy global smoothness, which is why our experiments use the theory to derive empirical proxies.
2. In line 234, δmin and δmax are unclear because only the last checkpoint of the model is published, so why would one care about all the δ during the training phase when only δT is relevant?
A: We agree that δT is what matters for the deployed model, and our single-checkpoint bounds (Eq. 12 at t=Tmax) address it directly. These are also the only bounds an attacker can use, since they only have access to the final checkpoint.
The trajectory bounds are meant for the defender. The model developer has the full training trajectory, and hence CSL and CSG, at no extra cost. A small upper bound on δmin flags a sample that came close to the decision boundary during training, typically a hard or memorized sample, which the developer can audit and mitigate before release. The lower bound on δmax certifies the best margin the sample reached. Moreover, when the released checkpoint is chosen by early stopping, δmin and δmax bracket the distance at whichever checkpoint is selected. We will clarify this in the revision.
3. For equation 13, why is this the interesting direction of the inequalities when it seems more interesting to prove an upper bound to δmax and a lower bound to δmin?
A: We agree that the reviewer's directions would be more informative, since they would bound every δt, including δTmax. The current direction is forced by the fact that CSL and CSG are averages over the trajectory. To average the per-iteration upper bound, δt must be replaced by a quantity independent of t, and only δt≥δmin keeps the inequality valid. An upper bound on δmax would need control of every individual checkpoint, which an average cannot provide: one checkpoint with a vanishing gradient can make δt arbitrarily large while barely changing CSG. The same argument limits the lower bound to δmax. In addition, nearly every sample is misclassified at initialization, so δmin≈0 and a lower bound on it would be vacuous over the full trajectory.
The current bounds are still meaningful for the defender, as described above. We will add this as a remark and as a limitation of our work.
4. I don't agree with Theorem 2's interpretation because while δmin is the weakest point during training, the deployed model will have δT≥δmin so the upper bound may not apply to δT, and similarly, the lower bound may not apply to δT since δT≤δmax.
A: The reviewer is correct that the trajectory-based bounds for δmin and δmax do not apply to δT. The bounds for δT are our Single-Checkpoint bounds in Theorem 2 (Eq. 12), which serve both the defender and the attacker. Since the attacker only has access to the final checkpoint, these are the only bounds they can compute: the lower bound provides safety guarantees for the defender, and the upper bound provides attack-budget guarantees for the attacker. The trajectory-based bounds can only be computed by the defender. We will update the interpretation of Theorem 2 to clarify this.