Attacks on Online Learners: a Teacher-Student Analysis

Machine learning models are famously vulnerable to adversarial attacks: small ad-hoc perturbations of the data that can catastrophically alter the model predictions. While a large literature has studied the case of test-time attacks on pre-trained models, the important case of attacks in an online learning setting has received little attention so far. In this work, we use a control-theoretical perspective to study the scenario where an attacker may perturb data labels to manipulate the learning dynamics of an online learner. We perform a theoretical analysis of the problem in a teacher-student setup, considering different attack strategies, and obtaining analytical results for the steady state of simple linear learners. These results enable us to prove that a discontinuous transition in the learner's accuracy occurs when the attack strength exceeds a critical threshold. We then study empirically attacks on learners with complex architectures using real data, confirming the insights of our theoretical analysis. Our findings show that greedy attacks can be extremely efficient, especially when data stream in small batches.

Paper

References (33)

Scroll for more · 21 remaining

Similar papers

Peer review

Reviewer MYtD5/10 · confidence 3/52023-07-06

Summary

This paper sheds light on the vulnerabilities of online learners to adversarial attacks and provides insights into the manipulation of learning dynamics. The findings highlight the need for robust defenses against such attacks and contribute to the growing body of research on data poisoning attacks in machine learning.

Strengths

1. The paper provides a theoretical analysis of online data poisoning in the teacher-student setup, which is a popular framework for studying machine learning models in a controlled way. 2. The paper compares different attack strategies and finds that properly calibrated greedy attacks can be as effective as attacks with full knowledge of the incoming data stream. 3. The authors empirically study online data poisoning on real datasets, such as MNIST and CIFAR10, using various architectures like LeNet, ResNet, and VGG.

Weaknesses

1. This paper primarily focuses on linear regression models and does not extensively explore the impact of adversarial attacks on other types of machine learning models. 2. This paper does not provide specific examples of real-world scenarios where adversarial attacks on online learners can have significant consequences. 3. This paper lacks extensively discuss potential defense mechanisms or countermeasures to mitigate the impact of these attacks. 4. This paper does not thoroughly compare its findings with existing literature on attacks on online learners. 5. The practicality studied threat model should be discussed with more details.

Questions

1. How do different attack strategies compare in terms of efficacy? 2. Are attacks subject to time constraints and do they involve transient dynamics? 3. Does the analysis take into account temporal correlations within the data stream? 4. The model's robustness improves or decreases of over-parametrization?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

Missing limitation and negative societal impact discussion in this paper.

Reviewer 7DCo5/10 · confidence 4/52023-07-06

Summary

Data poisoning attacks have been extensively studied in the offline setting, while poisoning attacks in the online setting have received little attention. In the online setting, the attacker is forced to craft attacks exploiting the data stream, taking into account the state of the model and the possible future data stream to divert the model. This work aims to study theoretically and empirically the learning dynamics of an online learner subject to a poisoning attack in the white-box setting, i.e., the attacker knows the architecture and the current parameters of the target model. The problem of finding the best attack policy is formalized as a stochastic optimal control problem and is theoretically studied in the teacher-student setup. From that, the authors were able to obtain results about the steady state of an attacked linear regression model, i.e., the model weights, the optimal action and the distance of the target model depending on the cost of the attack, supposing both infinite and finite size batches. Furthermore, the work shows analytically that a greedy strategy can perform quite optimally. The experimental evaluation confirms the theoretical results on synthetic data and explores the dynamics of online data poisoning attacks on real data and non-linear predictors, showing that the theoretical findings almost hold in more complex scenarios.

Strengths

This paper makes a valuable contribution to the literature on data poisoning attacks, where there has been limited work in online settings. In particular: - The idea of analyzing the attacker’s control problem from the perspective of the teacher-student framework represents the strong main novelty of the paper, since the formulation of the attack problem as an optimal control problem has been discussed before. The analysis allows the authors to derive interesting results about the dynamics of the learning process of the model and the steady state of a linear regressor that have not previously been discussed in the literature. It is interesting to see that above a certain attack strength the accuracy of the attacked model may drop dramatically and perturbations above the threshold may not be sufficient to realize the attack. - The set of considered attack strategies sufficiently covers the possible options available to the attacker. The idea of showing that a greedy attack can be as effective as an attack with full knowledge is valuable. - The theoretical results are confirmed empirically and punctually. Moreover, the experimental results on non-linear architectures and real data suggest that the theoretical results may also hold for more complex scenarios, that’s an interesting finding. - The derivation of the theoretical results is explained in the appendix and the well-organized code to reproduce the experimental results is available (even though more instructions to reproduce the results may be added, like the Python version used by the authors to run the experiments).

Weaknesses

- The theoretical analysis is limited to consider linear regression as the student model, the standard normal distribution as input distribution and an infinite time horizon, so the theoretical results are a bit limited. However, it is recognized by the authors that more general results are missing because deriving them in a more general scenario is hard. Furthermore, the authors tried to derive some intuitions about the results in a more complex scenario through the experimental evaluation. Covering the theoretical analysis in a more complex setting and with a finite time horizon is an interesting future work to explore. - It is not clearly motivated why the attacker can only modify the labels instead of the entire inputs and why the attacker should be able to apply the same perturbation to the entire batch (which turns out to be a very strong assumption). Data poisoning attackers should assume little control over the data to be considered practical [P1, P2]. - The presentation of the theoretical results is a bit confusing since it is mixed with the presentation of part of the theoretical results. - The presentation of the attack strategies and their implementation lack details that cannot be found in the appendix. See the “Questions” section for more details. - From the paper, it is not evident what the take-home message is and what suggestions are provided to improve robustness in stream pipelines or make attackers more stealthy. The paper presents numerous useful analyses and extensive experimental settings but lacks a "Take-home message/Discussion" section that summarizes the overall findings and suggestions. While the authors have reported some of these in the conclusion section, I believe they can be further expanded with higher priority/attention. [P1] Back to the Drawing Board: A Critical Evaluation of Poisoning Attacks on Federated Learning. In IEEE Symposium on Security and Privacy (2022). [P2] Wild Patterns Reloaded: A Survey of Machine Learning Security against Training Data Poisoning. ACM Computing Surveys (2022).

Questions

Major questions: - Why are the clairvoyant and RL attacks infeasible to deploy for complex models? Is it due to the time required for the calibration? This is not very clear from the text. - In Eq. 1, please clarify why it should be important to model an attacker that “under/overshoots”. Why can the attacker obtain a perturbed label that resides outside the interval between the clean and target labels? - I think that more details about the RL attack are needed. How is the problem of finding the optimal sequence of actions mapped to the RL problem? Which are the reasons for choosing to utilize TD3 (and a deep learning policy in general) instead of other RL approaches? - Which is the utility of gamma^tilde in the objective function minimized by the greedy attack? Is it used to address the fact that the attacker should not try the most effective attack at the subsequent time step, but she has to take into account all the future streams of instances? How does the calibration of gamma^tilde work? I invite the authors to add more details about this. - I suggest that the authors add a "Take-home-messages/Discussion” section that summarizes the theoretical and empirical results and highlights how the results can be useful for designing a poisoning attack or defending against it. - The authors should clarify their threat model in the paper and where it may apply. Minor questions: - Are the results of Figures 1-B and 1-C obtained from the experimental evaluation? If yes, the setting used for obtaining these results should be indicated in the text. I may have missed it, so I kindly ask the authors to point me out the lines containing this information if it is included in the paper. - Should g^nef be at the place of one of the g^pre in eq. 16? Eventually, I kindly ask the authors to fix the links to the equations that appear in the text, since they don’t work at the moment.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The authors have briefly discussed some limitations of their work in the conclusion section and they have clearly stated that extending their analysis to non-linear architectures and different probability distributions of the inputs is challenging. However, I think that a discussion about the limitations of the threat model considered is missing and I invite the authors to cover this point, by discussing, for example, why the attacker should only modify the labels of the instances. Moreover, I ask the authors to include a “take-on-messages/discussion” section to summarize and clarify the obtained results and their implications.

Reviewer 61gD3/10 · confidence 3/52023-07-10

Summary

This paper analyzes the robustness of online learners when the labels of the received data are manipulated. In particular, it analyzes a student-teacher online learning problem where the attacker poisons the labels provided by the teacher before feeding the student the labeled batch. The setup is analyzed both theoretically and experimentally where with the increase of the attack budget, the performance of the learner degrades significantly.

Strengths

The main strengths of this work are: (1) The paper is well-written and the figures are polished. (2) The setup is clearly explained and the theoretical analysis on synthetic data support the results.

Weaknesses

There are few weaknesses that I hope to be addressed before getting this paper accepted: (1) The paper is missing a strong practical motivational example from real world scenarios when attackers have access to manipulate the *labels* before feeding the labeled batches to the learner. (2) While the proposed problem setup studies an online learning scheme, the attacker is allowed for an unbounded computational budget to manipulate the labels. However, in online learning, the environment can reveal new batches *irrespective* of how efficient the attacker is in manipulating the previously revealed batch. Note that this assumption (given unlimited computation to the attacker) is not realistic and it questions that effectiveness of the proposed attack. (3) Experimental evaluation are conducted on very small scale problems and datasets. I would expect evaluating the results on online learning datasets such as CLOC [A] and Firehose [B]. (4) The analyzed attackers should be evaluated when a defense mechanism is presented. For instance, would a learner upon convergence still update all model parameters on each received data? What happens if the learner stored labeled examples from the stream before the poisoning phase starts and update on mixed batches? [A] Online Continual Learning with Natural Distribution Shifts: An Empirical Study with Visual Data, ICCV 2021. [B] Drinking from a Firehose: Continual Learning with Web-scale Natural Language, 2020.

Questions

In addition to the points raised in the weaknesses part, I have the following question:\ - To guarantee that the attack is imperceptible (in case of perturbing the input), the attacker has to produce the perturbations under budget constraints (e.g. $\|\delta\|_\infty \leq \frac{8}{255}$). How is this constraint translated to the proposed setup? - The perturbed label in Equation (1) could produce a real number for integer valued labels. Couldn't this be a simple check for the learner to reject such samples from training on them?

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors have not addressed the limitations of this work such as the practicality of the proposed setup, the limited experimental setup (problem and dataset scale) along with how expensive is it to conduct such attacks.

Reviewer cwDs6/10 · confidence 2/52023-07-19

Summary

The authors conduct theoretical analysis and empirical evaluation of poisoning attacks in the online learning setting, where the attacker can intervene in the labels of sequentially provided data. As a result of the theoretical analysis, the authors show that the strength of the attack becomes discontinuously stronger as the batch size asymptotically approaches infinity. It is also shown that even a greedy attack can be as strong as that in the clairvoyance setting under an appropriate condition.

Strengths

It introduces a new problem setup for poisoning, and although the analysis is performed in a very simple setting, it leads to interesting results. It is interesting that a properly designed greedy attack is as effective as an attack in the clairvoyant setting. Theoretical analysis utilizing optimal control theory is not very common, and it is useful to introduce such methods.

Weaknesses

There is little discussion of threat models and attacker models when viewed as a security issue. There is no disagreement that poisoning in an online setting is a serious threat, but the paper does not discuss where the significance of considering this type of dirty label attack lies and whether it could be a significant risk. In particular, I wonder if there is any point in dealing with a setting where the batch size is infinite in an online setting.

Questions

In equation 4, is a parameter to adjust the balance between g^per and g^nef necessary? In equation 6, when d_mu(C,P)=1, it cannot be said that the student's prediction and the target's prediction coincide since they merely coincide in their expectations.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The subject is an attack on a learning system and it can affect society in a malicious way, but it is obvious that the research interest is in the theoretical aspect of the behavior of the system under attack.

Reviewer rUtu6/10 · confidence 3/52023-07-21

Summary

The paper provide an analysis of steady state of linear learners trained via SGD under online setting, where the found a phase transition in terms of attack strength. Some experiments are also conducted to demonstrate the insight from the theory.

Strengths

The formulation of attacking learning model under online setting is clean and intuitive. The theory is also illuminating. The insight that model performance will sharply decrease if attach strength is higher than a threshold is an interesting observation. In addition, the theory is supported by experiments.

Weaknesses

Overall my feeling is both empirically and theoretically, the paper is not strong. Theory side, the main concern I have on the work is that the problem might be oversimplified, thus it is not sure how much the insight can carry over to more realistic settings. On empirical side, there are quite a few simplifications in the setting analyzed, e.g., linear model, SGD only training. While I agree theory should start simple, it would be more convincing to see experiments on more realistic settings, i.e., to test if the insights from the simple model can hold for more practical settings, even if they do not hold, it would be still interesting to see the results.

Questions

1. If the training algorithm is not SGD, instead it is, say Adam, how would this change the analysis result conceptually? In addition, in experiment would it change anything empirically? 2. In Fig. 3, reinforcement learning and clairvoyant methods are not compared for neural nets due to their complications. I can understand clairvoyant might be infeasible due to high nonconvexity of the problem, but why RL methods are not feasible? It seems like the online adversarial attack could be a reinforcement learning problem by nature.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Limitation are sufficiently discussed.

Authorsrebuttal2023-08-15

We are at the reviewers' disposal for any further questions

Since the discussion phase is about to end, we are writing to kindly ask if the reviewers have any additional comments regarding our response. We are at their disposal for any further questions. In addition, if our new experiments address the reviewers' concerns, we would like to kindly ask if the reviewers could reconsider their scores.

Reviewer 7DCo2023-08-21

Thanks to the authors for their clear response to my concerns. Similarly to other reviewers, I initially held reservations regarding the considered threat model and the current scope of the proposed work. However, despite these reservations, I have chosen not to lower my original score. This decision is primarily influenced by the significant and noteworthy experimental findings that the authors have presented in their rebuttal. In conclusion, the paper will require several clarifications in its final iteration. These clarifications should encompass aspects such as the delineation of the threat model, underlying assumptions and limitations, and the rationale behind the chosen attack strategy. As pointed out by all reviewers, these elements are pivotal to enhancing the clarity and rigor of the paper. Furthermore, a more seamless integration of the outcomes presented in the main body of the paper with those provided in the appendix is advised for achieving a coherent narrative.

Authorsrebuttal2023-08-21

We thank the reviewer for the response and the positive evaluation of our supplementary findings in the rebuttal. We are pleased that our response effectively addressed the reviewer's concerns. Overall, we believe that our submission has significantly benefited from the reviewers' feedback. In the final version of the manuscript, we will incorporate the clarifications mentioned by the reviewer, as we discussed in our rebuttal.

Reviewer 61gD2023-08-21

Dear Authors I would like to thank you for the efforts put into writing the rebuttal. However, several concerns of mine are still unresolved. While I understand that the nature of this work is to establish baseline behavior's of online learners, it is still required to show the importance of this problem when the learner is somewhat smart. That is, one should consider *at least* simple defenses and show that such approaches fail in defending against the proposed attack. Second, since the problem setup tackles online learning, I am not sure about the validity of the results when considering stationary and small scale distributions such as CIFAR10 and MNIST. Hence, one should include distribution-changing datasets that are designed specifically for online learning (such as the one mentioned in the review) and study the problem there instead of small scale datasets. Third and regarding the unbounded computational budget: while I appreciate this extension, one should study the effect of making $\rho$ a function of how expensive the developed attack. Further, it is important to conduct the empirical experiments under this setting as well. That being said, I think that this work is not ready to be accepted and hence I will maintain my original score.

Authorsrebuttal2023-08-21

We thank the Reviewer for the response. We are happy that our rebuttal addressed, at least partially, the Reviewer's concerns. We would kindly request the Reviewer to consider the following points before reaching their final decision: * **Simple defense mechanisms**. The design and effectiveness of defense strategies greatly rely on the specific attack scenario. For instance, in the context of federated learning, anomaly detection filters might be employed on weight updates by the central server. Conversely, in centralized learning, data may undergo scrutiny, such as feasible-set projection, to ensure label integrity (an example is to control that labels are integers, as discussed before). While we recognize the importance of addressing these instances, the study of defense mechanisms falls outside the scope of our current work. Our primary focus is on offering a broader understanding of label attacks on online learners, serving as a foundational framework for future explorations. * **Non-stationary datasets**. We appreciate the Reviewer's suggestion. We would like to point out that the experiments we carried out are very well aligned with the related literature [1, 2, 3, 4]. We will carefully consider non-stationary distributions as a possible direction for future work, as we acknowledge the relevance of the proposed analysis. * **$\rho$ experiments**. We would like to point out that, in our rebuttal, we did run empirical experiments using MNIST (see the attached file). Due to time constraints during the rebuttal phase, we restricted our analysis to this dataset and a simple logistic regression model. We are currently running the $\rho$ experiments for all empirical scenarios outlined in our draft, and we will include the results in the final version of the manuscript. Once again, we thank you for your valuable feedback. [1] Pang, T. et al. Accumulative Poisoning Attacks on Real-time Data (NeurIPS, 2021). [2] Lee, S. et al. Continual Learning in the Teacher-Student Setup: Impact of Task Similarity (ICML, 2021). [3] Zhang, X. et al. Online Data Poisoning Attacks (L4DC, 2020). [4] Wang, Y. and Chaudhuri, K. Data Poisoning Attacks against Online Learning (arXiv, 2018).

Reviewer cwDs2023-08-22

Dear Authors Thank you for your responses to the reviewer's questions and comments. After reading the responses, some of my questions are resolved. As the other reviewers pointed out, this work requires several clarifications reflecting the reviewers' comments for publication while I would like to keep the current score since I believe this manuscript contains an interesting approach to attacks on online learning.

Authorsrebuttal2023-08-22

We thank the Reviewer for the positive evaluation of our submission. We are glad the Reviewer finds our work important for online adversarial attacks, and that we resolved part of the Reviewer's questions. Although the official discussion phase has concluded, we remain available to address any additional questions the Reviewer might have. We will of course incorporate in the final version of the manuscript all the clarifications that we provided in our rebuttal. Thank you again for your valuable feedback.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC