acceptodds
Withdrawn from ICLR 2027

When and How Often to Defend? On the Overlooked Role of Learning Rate Schedules in Poisoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Data-poisoning defenses are typically applied throughout training, even though they can substantially reduce model accuracy. In this work, we ask: when and how frequently should a defense be applied? We highlight the overlooked role of the learning-rate schedule and prove that the effect of a defense at a given step scales with the squared learning rate . Defending therefore matters most within the window containing the largest cumulative mass. We further relate defense decisions to the current gradient-state information and classify training-time defenses into three regimes: state independent, state dependent, and mediator. We evaluate these regimes against the seminal Witches' Brew attack on CIFAR-10, CIFAR-100, and Tiny ImageNet, using a DP-SGD-style defense applied to the aggregated gradient. Our experiments show that attack success rises primarily around the first learning-rate drop despite a stable poisoning frequency. When the defense is restricted to the window carrying most of the mass, gradient-state information becomes valuable: a state-dependent policy matches or outperforms a random policy at a lower activation rate. Most importantly, activating the defense at only a small fraction of training steps, well distributed across the schedule, fully prevents the attack while only slightly reducing clean accuracy, yielding a better security/accuracy trade-off. Effective defense therefore requires far fewer interventions than commonly assumed, and where theses interventions fall on the learning-rate schedule matters as much as their frequency.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Poisoning, Security, Defense, Training-time attacks, Learning rate schedule, Witche's Brew

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_JVQpOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies when and how often to apply a training-time defense against clean-label data poisoning. It uses a bound on the one-step change in adversarial loss to motivate concentrating defense interventions in a squared-mass learning-rate (SMLR) window, where much of the schedule’s cumulative squared learning rate lies. The paper also classifies defense policies according to whether their activation decisions use the current gradient state. Using a clipping-and-noising defense against the Witches’ Brew attack, it compares always-on, random, oracle, and window-based policies across several image-classification settings. The experiments report that sparse interventions, particularly within the SMLR window, can often prevent the attack with less clean-accuracy loss than applying the defense at every training step.

Critical strengths

Focusing on the learning-rate schedule and the timing of defense interventions offers a novel perspective on training-time defenses.

Critical weaknesses

The writing is poor and frequently confusing. Many symbols appear without formal definitions at the point where they are used. For example, at the bottom of page 3, the authors write Ladv(θ):=L(F(xt,θ),yadv)L_{\mathrm{adv}}(\theta):=L(F(x_t,\theta),y_{\mathrm{adv}}) without defining FF. The roles of xtx_t, (xi,yi)(x_i,y_i), and yadvy_{\mathrm{adv}} are also not clearly specified there. In Eq. (2), B(Δ,θk)B(\Delta,\theta_k) depends on Δ\Delta, but this dependence is not explicit on the right-hand side. Although the text calls Δ\Delta a fixed poison, it does not formally specify whether it denotes perturbations to individual samples, the resulting poisoned dataset, or how it enters L(θ)L(\theta).

My second and more serious concern is the soundness and logic of the paper’s central finding. The focus on ∑k=1Kηk2\sum_{k=1}^{K}\eta_k^2 is motivated by Proposition 1: its left-hand side measures the expected one-step change in adversarial loss, while a positive term proportional to ηk2\eta_k^2 appears on the right-hand side. It is reasonable that defending at different training steps may have different effects, but what this proposition actually says about those effects remains unclear. I take the authors’ intended intuition to be that defending at steps with larger ηk2\eta_k^2 makes Ladv(θK)L_{\mathrm{adv}}(\theta_K) larger. However, Eq. (5) accumulates only the positive term from an upper bound. Making that term larger merely loosens the bound. It does not guarantee that the actual adversarial loss increases or that the attack fails. Moreover, the first two terms in Proposition 1 also depend on the learning rate and may change under different defense policies. The authors do not explain why these terms can be set aside while the third term determines the scheduling rule. Finally, even if the analysis established a higher adversarial loss, it is unclear why the learner should prioritize it over its own objective, L(θ)L(\theta). Increasing adversarial loss is not the same as minimizing the learner’s training loss. Due to the poor explanation, it is not even clear when Eq. (5) is first presented whether the authors intend the squared learning-rate mass to be increased or decreased. As a result, the theoretical justification for the paper’s central claim is fundamentally unsound.

Questions for authors

See weakness.

Reviewer_fDR7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies the relationship between the effect of a poisoning defense and the learning rate schedule. It identifies that the effect of the defense at a given step scales with the squared learning rate. Consequently, the authors propose to apply the defense only within the training window containing the largest cumulative squared learning rate. The findings are evaluated with a DP-SGD defense against Witches’ Brew poisoning attack on common benchmarks.

Critical strengths
  1. The paper studies an interesting question and is practically motivated.

  2. The theory is encouraging, though it is an upper bound.

  3. Empirical analysis shows defending only for certain steps closes the gap with clean accuracy, which adds value.

Critical weaknesses
  • The effectiveness of a defense is expected to be proportional to the learning rate. The exact relationship is novel. However, I have a couple of concerns related to the theory.
  1. I think the derived upper bound in eq 4 is loose and doesn't clearly establish the actual defense contribution. For instance, if the adversarial loss is affine, which is L-Lipschitz, but the zero mean noise has exactly zero effect on the expected adversarial loss. Therefore, the looseness in the bound makes me question the reliability of the result.
  2. Another concern is that the theory assumes gradient descent, but the experiments use Nesterov momentum and weight decay. So, it is not clear whether the shown dependence on the learning rate still holds. A clarification on this would be helpful.
  3. In line 699, ∣∣g~k∣∣=C||\tilde{g}_k||=C is assumed which wouldn’t hold when ∣∣gk∣∣<C||g_k|| < C. It should either be included as an assumption in the proposition or a more precise analysis should be done.
  • I find the empirical evaluation to be limited. The evaluation includes only one poisoning attack (Witches’ Brew) and one defense (DP-SGD) under only linf perturbations.

  • The proposed taxonomy in Section 3.2 is not clearly disjoint. The identity defense decision is also independent of the current gradient state. Therefore, the proposed mediator category needs a more precise distinction from regime A.

Questions for authors

I have some questions for the authors. Please also check Weaknesses.

  1. In the experiments, the results are averaged over the last 10 epochs (line 323). This measures across correlated checkpoints but not across independent runs. So, the statistical significance of results like 0 attack success rate is unclear.

  2. Poisoned samples are constructed using a pretrained model, with no reported adaptation to randomized defense or its schedule. Does the defense remain effective when poison construction explicitly accounts for the proposed defense schedule? One could test this using adaptive Witches' Brew attack.

  3. Since the evaluations are based on fixed poison samples, claims of full prevention or an optimal security and accuracy tradeoff require substantially stronger evidence. Evaluation of one attack and one gradient transformation also limits the broader conclusions. Therefore, I suggest including additional attacks like MetaPoison [1], Bullseye polytope [2], Sleeper agent [3], etc.

  4. How sensitive are the results to the window threshold, clipping norm, and noise multiplier? Can a continuously applied defense with a smaller noise multiplier achieve a comparable tradeoff?

[1] Huang et al. MetaPoison: Practical General-purpose Clean-label Data Poisoning. NeurIPS 2020.

[2] Aghakhani et al. Bullseye polytope: A scalable clean-label poisoning attack with improved transferability. S&P 2021.

[3] Souri et al. Sleeper Agent: Scalable Hidden Trigger Backdoors for Neural Networks Trained from Scratch. NeurIPS 2022.

Reviewer_S2stOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript attempts to answer the questions of when / how frequently / and how to defend effectively against data poisoning attacks to achieve the best ASR-Utility tradeoff. In order to defend the gradient matching attack, the paper reports that it is only necessary to implement defensive steps at selected steps within a window where the learning rates are substantial. The manuscript evaluated the proposed method against the gradient matching attack on the CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets.

Critical strengths

None

Critical weaknesses
  1. This paper attempts to answer three questions: When to defend, how frequently to defend, and how to defend effectively against data poisoning. However, the empirical evaluation is restricted to the gradient-matching attack and an existing DP-based defense method. It is unclear whether the findings can extend to other data poisoning attacks or defenses. In its current state, it feels as if the manuscript does not answer the three questions itself proposed.

  2. There are various statements / sections in the manuscript that appear incomplete or unjustified. For example, Section 3.2 claims but does not establish why the the ASR-utility trade-off motivates examine the gradient states. As a result, the authors' decision to use oracle models as a baseline, instead of defense methods introduced in prior works, does not completely justified. Additionally, the rationale for the selected activation ranges and threshold could be clearer. Sensitivity analyses and comparisons at matched intervention budgets would help establish how much the conclusions depend on these choices.

Questions for authors
  • Could the authors please clarify how could selecting steps with the largest learning rates lead to a better ASR-Utility tradeoff? It seems like the high learning rates are amplifying the defense's effect disrupting useful learning as well?

  • In Section 4.4 the manuscript claims that the Sparse random interventions provide the most reliable ASR-Utility trade-off. However, from Figure 5, it looks like Oracle 2 is outperforming random-50% under both step scheduling and restart scheduling? Could the authors please provide the final accuracy from Figure 5 in a table, similar to Table 5?

  • Could the authors please provide some details on the gradient matching attack? How was the target image selected? Was there a single or multiple target images? The authors may want to report error margins from multiple runs with different hyperparameters (such as target image and poisoning budget) since its they can impact the final result to a great extent.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.