BAN: Detecting Backdoors Activated by Adversarial Neuron Noise

Backdoor attacks on deep learning represent a recent threat that has gained significant attention in the research community. Backdoor defenses are mainly based on backdoor inversion, which has been shown to be generic, model-agnostic, and applicable to practical threat scenarios. State-of-the-art backdoor inversion recovers a mask in the feature space to locate prominent backdoor features, where benign and backdoor features can be disentangled. However, it suffers from high computational overhead, and we also find that it overly relies on prominent backdoor features that are highly distinguishable from benign features. To tackle these shortcomings, this paper improves backdoor feature inversion for backdoor detection by incorporating extra neuron activation information. In particular, we adversarially increase the loss of backdoored models with respect to weights to activate the backdoor effect, based on which we can easily differentiate backdoored and clean models. Experimental results demonstrate our defense, BAN, is 1.37$\times$ (on CIFAR-10) and 5.11$\times$ (on ImageNet200) more efficient with an average 9.99\% higher detect success rate than the state-of-the-art defense BTI-DBF. Our code and trained models are publicly available at~\url{https://github.com/xiaoyunxxy/ban}.

Paper

Similar papers

Peer review

Reviewer cerE7/10 · confidence 4/52024-06-14

Summary

The paper proposes a novel detection and defense method for backdoored models. The new method is motivated by the finding that existing state-of-the-art trigger-inversion methods, like BTI-DBF, rely on strong backdoor features, which might not always be present, such as for BadNet-type triggers. As a solution, the paper proposes BAN (Backdoors activated by Adversarial neuron Noise), a novel detection method based on the findings that models with backdoors are more sensitive to adversarial noise than benign ones, allowing the identification of neurons responsible for the backdoor function. Practically, BAN computes neuron-specific noise inspired by adversarial attacks like PGD to maximize the model's classification loss on clean and unseen data. The method tries to decouple benign and backdoor features by optimizing a sparse neuron mask on the loss behavior of the model. Intuitively, a backdoored model tends to predict the target class label for clean inputs and the adversarial perturbed neurons, whereas a benign model shows fewer misclassifications under comparable neuron noise. Evaluated against common backdoor attacks and defense methods, BAN demonstrates improvements to existing approaches and robustness against adaptive attacks.

Strengths

- The proposed defense method adds an interesting aspect to existing backdoor defense methods by improving the feature decoupling between benign and backdoor features. Whereas the decoupling optimization is conceptionally simple (which is not bad at all), it demonstrates strong results on the evaluation benchmarks and beats existing detection and defense methods markedly. - The paper is well-written and easy to follow. Most parts of the proposed method is sufficiently placed in existing research, and the open research problems and the proposed solution are clearly presented and described. I enjoyed reading the paper. - The evaluation is comprehensive and compares the proposed method against numerous attacks and defenses. It also includes the all-to-all setting, which is often ignored in literature. Evaluations of adaptive attacks underline the effectiveness and robustness of the approach. - Empirical solutions and theoretical considerations sufficiently support all claims in the paper.

Weaknesses

- The evaluation focuses on common (comparably shallow) CNN architectures. However, given that ViT also plays an increasing role in image classification, showing the effectiveness of the approach on ViT architectures in contrast to traditional CNNs would further support the efficiency claims of the method. - The evaluation further focuses on dirty label attacks, i.e., attacks that change the label of a training example. However, the method's efficiency on clean-label attacks is not demonstrated. Including 1-2 clean label attacks in the evaluation would further strengthen the results. - Some contributions of the paper should be stated more clearly and set apart from existing approaches in literature. For example, the feature decoupling process described in 3.3 is quite similar to the method by Xu et al. [1], and it is unclear to me (by only reading this paper) what exactly distinguishes the proposed method from the one in the existing literature. Small remarks: - Some captions are missing details. For example, the caption of Table 5 does not clearly state the investigated setting. Which dataset and model architecture are used here? - The font in the figures could be increased. Some legends and texts are hard to read without zooming in. - There is a typo in line 142: "experiemnts" -> experiments. - Table 7 seems like a sensitivity analysis (measuring the sensitivity of the method to the hyperparameter selection) instead of an ablation study (impact of deactivating certain parts of the method). [1] Xu et al. "Towards Reliable and Efficient Backdoor Trigger Inversion via Decoupling Benign Features". ICLR 2024

Questions

- L303: "FeatureRE is designed for all-to-one attacks, so we use target label 0 here for FeatureRE." -> Does it make sense to include an approach designed for all-to-one attacks in an all-to-all setting? The comparison here with the other approaches seems misleading. - Table 4: What are the BA (benign accuracy) and ASR (attack success rate) of models trained only on clean data without any poisoned data samples? It would be interesting to compare the defended backdoored models to a clean model. Since ImageNet (and its subsets) are complex datasets to learn, there will also be misclassifications that might, by chance, lead to the prediction of target class labels (even without any backdoor integration). Given ASR of the defended models, it would be interesting to see a baseline ASR (achieved only by random misclassifications) to assess if the remaining ASR after applying the defense methods is due to remaining backdoor behavior or simple random model behavior. - The proposed method is based on the assumption that under neuron noise, a backdoored model tends to predict the target class label. In contrast, the predicted class labels for the benign model are rather equally distributed. This makes sense for datasets with clearly distinguishable classes, e.g., CIFAR10. Let us assume that there exist two classes in the dataset that are semantically very similar, e.g., two visually similar (but still different) dog breeds among other clearly separable classes (bug, plane, etc.). Let us further assume that the backdoor target is to change the label from the one dog class to the other one. If we now apply the neuron noise approach to a benign and a backdoored model, can we still assume that only the backdoored model will tend to predict the other dog class (the target class) under neuron noise? Or will the benign model under neuron noise behave similarly, since the other dog class is probably close in the feature space, and adding noise to the neuron activations could lead to predictions for the second dog class (which happens to be also the backdoor target)? Phrased differently, can we assume the proposed BAN method also works reliably on datasets with semantically similar classes, if the backdoor targets label changes from one class to the other?

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

Limitations are sufficiently discussed in Appx. A. Since the paper proposes a novel defense method for backdoor attacks, no negative societal impact is expected.

Reviewer cerE2024-08-08

I thank the authors for the detailed rebuttal and additional insights. All my questions and remarks were addressed. After reading all other reviews and author responses, I decided to increase my rating.

Authorsrebuttal2024-08-08

Thank you

Thank you for your quick reply and raising the score. We are happy that we have addressed your concerns

Reviewer Eja26/10 · confidence 4/52024-06-26

Summary

This paper provides an in-depth analysis of the SOTA trigger inversion-based backdoor defenses and finds that they suffer from high computational overhead and rely on prominent backdoor features. The authors tackle the challenges based on the previous findings on adversarial noise incorporating activation information and propose to improve the backdoor feature inversion for backdoor detection. The experiments show its efficiency and effectiveness compared to the SOTA methods.

Strengths

1. The proposed method is intuitively reasonable and empirically effective. Both backdoor detection and defense are carefully considered. 2. The paper writing is clear and well-structured. 3. The experiments are comprehensive.

Weaknesses

1. The motivations of high computational overhead and reliance on prominent backdoor features are not well-illustrated in the method section. Only the limitations of BTI-DBF are discussed. 2. The novelty is limited. The finding on neuron noise is previously proposed, and an additional mask regularizer based on BTI-DBF is a commonly used technique, which was used in NC as well. The idea of backdoor defense is similar to ANP, with the change in learning from the noise.

Questions

None.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

1. The reliance on prominent backdoor features is expected to be illustrated in detail, and the reason why the proposed method is able to overcome it with only a mask regularizer is unclear. Some case visualizations are expected. 2. The conclusion from Equation 7 that ignoring the second term is expected to be illustrated with evidence such as the separate loss values on it.

Reviewer U3ov7/10 · confidence 4/52024-07-08

Summary

This paper addresses the problem of efficient backdoor defense using the backdoor inversion approach. The authors leverage the past work “TOWARDS RELIABLE AND EFFICIENT BACKDOOR TRIGGER INVERSION VIA DECOUPLING BENIGN FEATURES” (BTI-DBF) which recovers a mask in the feature space to locate prominent backdoor features and decouples benign and backdoor features in the mask. The authors make contributions by improving the BTI-DBF method and by incorporating extra neuron activation information into their method called “Backdoors activated by Adversarial neuron Noise” or BAN.

Strengths

The strengths of the paper lie in (1) studying the trigger inversion-based detection methods (2) optimizing the performance of the BTI-DBF method by adding a regularizer to the loss function (3) incorporating adversarial activation noise into the BAN method

Weaknesses

The weaknesses of the paper lie in (1) missing explanations of the performance. For example, why is featureRE in Table 3 performing better than BAN? What would be the impact of a lambda value other than 0.5? Why is the Blend attack always better detected by BTI-DBF* than by BAN? (2) missing relationship between the adversarial activation noise approach and input/feature perturbation approach? How is the adversarial activation noise approach different from assuming Lipschitz continuity on the estimated input-output function?

Questions

The authors should answer the questions posed as weaknesses of the paper.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The limitations of the work lie in the accuracy and computational complexity of BAN. The authors describe the limitations in Appendix A and focus the limitations on the availability of clean (benign) samples and a slight decrease in the accuracy for benign inputs after fine-tuning. The authors state, “We also exploit the neuron noise to further design a simple yet effective defense for removing the backdoor, such that we build a complete defense framework.” The statement “we build a complete defense framework” is an overstatement in the paper.

Reviewer hJLG5/10 · confidence 2/52024-07-09

Summary

This paper focuses on the backdoor defense task. The proposed detection method includes neuron noise, which leverages the differing robustness between regular and backdoored neurons, and a feature decoupling process using a mask. Additionally, a backdoor defense method is proposed, which achieves improved efficiency and overall performance.

Strengths

The exhibited efficiency is impressive, and the motivation behind neuron noise is intriguing. The dimensions of conducted experiments are relatively thorough.

Weaknesses

The neuron noise approach appears reasonable and interesting, but the feature decoupling with a mask is a little questionable. Additionally, the study does not include enough baselines, such as fine pruning, which can obscure the distinction between backdoored and regular neurons.

Questions

1, I might have a misunderstanding, but my key question is that the paper emphasizes that previous methods overly rely on prominent backdoor features. However, the Feature Decoupling with Mask at the neuron level seems to rely even more on these prominent backdoor features. Some previous works, like fine pruning [1], show that neurons might not necessarily be decoupled. 2, Additionally, in Section 4.4, the paper claims that "the reason is that the backdoor features of SSDT are close to benign features in the feature space. It is difficult for other methods to distinguish between backdoor and benign features created by SSDT." However, the Feature Decoupling with Mask approach may not be effective in this context. 3, There is a minor issue that the experimental improvement does not seem significant, especially in Table 3 and Table 4. 4, Is the detection performance related to the pattern of the trigger or the training strength of the backdoor? 5, I'm curious about when a model architecture becomes less redundant for a given dataset. For instance, when we train DenseNet121 on CIFAR-10, the initial features are likely to be sparse. On the other hand, training an MLP model on CIFAR-10 makes feature decoupling more challenging. I wonder if using smaller models makes feature decoupling less effective. 6 I also believe that selecting the lambda value in Equation (5) is quite challenging and tricky. This is because it is essential to maintain performance and ensure it doesn't degrade significantly after applying the mask (1-m). I would like to change my score if the questions are well answered. [1] Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks

Rating

5

Confidence

2

Soundness

3

Presentation

2

Contribution

3

Limitations

The limitations have been discussed.

Reviewer yW417/10 · confidence 3/52024-07-12

Summary

The authors propose a novel technique for detecting backdoor attacks on neural networks by incorporating extra neuron activation information to reduce the overhead from prior backdoor feature inversion methods. The experimental results show a higher detection rate on the tested datasets when compared to the prior work.

Strengths

The paper is well written and easy to follow. The authors address a significant topic with the widespread adoption of neural networks for a wide variety of tasks. The experimental results affirm the design choices made by the authors in Section 3. Overall, it is a well written paper that addresses a significant area.

Weaknesses

1. A figure of the proposed method outlined in Sections 3.1 and 3.3 could be a helpful tool to visualize the proposed method. 2. The in figure text in Figures 2 and 3 are too small and hard to read. 3. Some tables use % and others don't when report accuracy, e.g. Table 2 and Table 3 with no % symbol for the BA columns. 4. I think a table or figure further emphasizing the proposed changes would greatly improve the strength of this paper.

Questions

1. What is the impact of $\lambda_2$ on the fine-tuning loss found in Equation 8?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes, in appendix A.

Reviewer yW412024-08-07

Thank you for the responses to all of my questions. I have read the rebuttal and do not have any further questions myself.

Reviewer Eja22024-08-08

Thanks for the author's efforts in rebuttal. I am satisfied with the discussion on the prominence of features and my concerns are addressed. I will raise my score.

Authorsrebuttal2024-08-08

Thank you

Thank you for your quick reply and raising the score. We are happy that we have addressed your concerns

Reviewer U3ov2024-08-11

I have read the rebuttal. I do not have additional comments assuming that the authors are going to make changes and add the extra references according to their rebuttal.

Authorsrebuttal2024-08-11

Thank you

Thank you for your reply. We are happy that we have addressed your concerns. We will revise our work according to the rebuttal.

Reviewer hJLG2024-08-11

I appreciate the authors' detailed rebuttal and the additional insights provided. Although most of my questions have been addressed, I still find the novelty somewhat unconvincing to warrant an increase in my rating. However, I agree with accepting the paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC