Are aligned neural networks adversarially aligned?

Large language models are now tuned to align with the goals of their creators, namely to be "helpful and harmless." These models should respond helpfully to user questions, but refuse to answer requests that could cause harm. However, adversarial users can construct inputs which circumvent attempts at alignment. In this work, we study adversarial alignment, and ask to what extent these models remain aligned when interacting with an adversarial user who constructs worst-case inputs (adversarial examples). These inputs are designed to cause the model to emit harmful content that would otherwise be prohibited. We show that existing NLP-based optimization attacks are insufficiently powerful to reliably attack aligned text models: even when current NLP-based attacks fail, we can find adversarial inputs with brute force. As a result, the failure of current attacks should not be seen as proof that aligned text models remain aligned under adversarial inputs. However the recent trend in large-scale ML models is multimodal models that allow users to provide images that influence the text that is generated. We show these models can be easily attacked, i.e., induced to perform arbitrary un-aligned behavior through adversarial perturbation of the input image. We conjecture that improved NLP attacks may demonstrate this same level of adversarial control over text-only models.

Paper

References (57)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer n3yx4/10 · confidence 5/52023-06-17

Summary

This paper investigates the alignment of large language models in being helpful and harmless. While current NLP attacks are not powerful enough to consistently attack text models, adversarial input images can lead multimodal models to emit harmful content.

Strengths

The paper focuses on studying the NLP attacks which might trigger adversarial behavior on aligned models. The research line seems novel.

Weaknesses

1. The paper's structure is disorganized, making it difficult for readers to follow. For example, the background section lacks a detailed explanation of existing attack methods and their comparison with the proposed approach. Additionally, there is repeated summarization of previous work throughout different sections, resulting in a confusing structure. Section 4.1 discusses prior attack methods, but important related works such as GBDA and ARCA are not introduced in Section 4.3. 2. The expression in lines 171-175 is unclear, and it is difficult to comprehend the intended message. It would be helpful to provide a more detailed explanation of the two challenges mentioned. Why does it require |S| iterations of entire neural network forward passes? Why can't the sentence-level embeddings be directly inputted into the toxicity classifier for classification? 3. The paper lacks sufficient details in explaining the experimental setup. For example, what does "StableVicuna" in Table 1 refer to? What value is chosen for K in line 216? Which vision model is used in line 319? 4. The explanation of how the paper utilizes a brute force method in lines 252-257 needs further clarification. The description provided is confusing, and a more detailed explanation would be beneficial. 5. TYPO: In line 258, "Tto" should be corrected to "To."

Questions

still concern about the writing issues since the weakness part, more concerns here, 1. Generally, it is not clear from the abstract what the contributions of this paper are. It is not until page 7 that the authors briefly explain what they are researching, and sections 1-5 contain a lot of writing that lacks meaning or is confusing. 2. The introduction is difficult to understand as it is unclear what is being discussed. What are "continuous-domain images," and what is being conveyed in lines 56-62? 3. What does the last sentence of background “we make ... often will not be” mean? 4. What does the name of Sec3 “Threat Model” refer to? 5. What is the “Attack Goal”? Line 152 – 158 do not bring any valuable information. Your definition is also confusing, and using substring detection for harmful text is not accurate because it is impossible to cover all possible harmful strings. 6. I don't understand 171-172. If discrete tokens cannot be optimized, then how are all current language models trained? I don't know what 171-175 is referring, either. 7. What is the “initial attack objective” in Line 182. 8. Sec 5.1 and 5.2 appear to be your contributions, but I couldn't understand them, including your analysis in 6.3 and 6.4. They are too difficult to comprehend.

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

see the weakness and question parts.

Reviewer 5kqR6/10 · confidence 5/52023-07-02

Summary

The paper investigates the robustness of aligned neural networks, such as language models, by optimizing adversarial examples that aim to elicit unwanted behaviors (e.g., toxicity). The paper applies existing SOTA NLP adversarial attacks and finds they are insufficient to attack aligned models. However, the paper reports that the adversary's job is easier with multi-modal models due to the continuous space of image manipulation. Manipulating the image would cause the model to output unwanted and harmful content.

Strengths

- The paper takes the first step towards synergizing adversarial ML research with the new attack vectors of chatbot and LLM security and alignment. I find this a great initiative to start a line of research to understand the similarities and differences between the two domains. The paper also constructs successful attacks against multimodal models. I find the paper generally well written (except for the need for some clarifications and more details). I appreciate the discussion that the absence of successful attacks should not be interpreted as proof that models are adversarially aligned.

Weaknesses

- Despite the previous strengths, the paper does not show significant results as the examined attacks are not as successful compared to unaligned models (e.g., GPT-2). With the absence of any successful attacks, I am not very convinced with the paper's contribution and, taking the paper's initiatives aside, how much it can help advance research in this domain. - Since the multimodal attacks show more potential, I would have appreciated deeper evaluation and discussion regarding this attack's potential. This would have increased the significance of the paper. For example, I am curious to know: how does the prompt (the user's query) impact the attack's success? How generalizable is the attack over different prompts given the same optimized image? How can the attack be stealthy (to include the threat model of malicious third-party data)? How would such attacks be different for other objectives (beyond toxicity)? How to produce more adversary-controlled attacks? I, of course, don't expect the paper to answer all these questions, but I believe more focus could have been put on evaluating these attacks. - For reproducibility and better clarity, I think the experiments in section 6.2 need to have more details (to describe the attack optimization), possibly in an appendix. - I was very confused by the experiment in sections 5.1 and 5.2, and I only started to get them after a couple of iterations. I was confused that 1) they are performed on GPT-2 (unaligned model), 2) the target token $q$ does not seem to be exclusively toxic, and 3) they don't follow the chatbot format explained in 4.1. I think this section needs to explain the motivation (as far as I understood, it shows the limitation of current attacks even on unaligned models and that it cannot find the prefix $p$ even though it always exists by construction). - Minor comment: line 151: I don't think the content filters are exclusive to the malicious third-party threat model. It seems that current systems (e.g., Bing) perform content filtering over users' input as well. - Minor mistakes: line 258: Tto line 159: an language model missing citation in the caption of table 3.

Questions

- Why is the attacks' success rate defined by word overlap? Is it possible to use a classifier?

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The paper addressed the ethical consideration and limitations.

Reviewer NFJs5/10 · confidence 4/52023-07-04

Summary

**Summary:** This paper discusses the alignment of neural networks, specifically language models, in the context of adversarial attacks. The authors explore the extent to which language models remain aligned even when interacting with adversarial users who construct worst-case inputs. They find that existing NLP-based adversarial attacks are insufficiently powerful to reliably attack aligned text models, but multimodal models can be caused to emit hateful and toxic content through adversarial perturbation of the input image. The authors hope that stronger NLP attacks will enable comprehensive robustness evaluations of aligned LLMs, which can help inform alignment researchers to develop improved model alignment techniques that remain reliable in adversarial environments. **Pros:** 1.The paper addresses an important and timely topic in the field of AI, which is the adversarial robustness of language models. 2.The authors provide a clear and concise explanation of the problem and their approach to studying it. 3.The paper presents empirical evidence to support their findings, which can be useful for other researchers in the field. **Cons:** [Concept misleading] The concepts overlap between harmness and adversary. Adversarial attack is a set of algorithms that adds adversarial perturbation to sentences cumulatively based on some discrete optimization techniques. While harmness is a property that reflects the generation quality in terms of ethics. [Insufficient survey] The authors choose the title focusing on adversarial attack, they’re actually talking about preventing harmness instead of general adversarial attack, so the title is too wide. There’re lots of defenses and attacks in the NLP field, which has not been discussed by the authors, only mentioned in related work, and they didn’t discuss the relevance of such works. [Limited novelty] This paper mainly discusses how to hack language models to produce harmful contents without their own designed methods. Overall, it’s an empirical analysis with limited baselines, models and benchmarks. [Writing issue] The authors use too much space to discuss the background on adversarial attack, over two pages. In fact, using one simple paragraph can cover all the basics in an adversarial attack. Besides, some parts of the paper are not necessary, for example, the Multimodal text-vision models in related works can be omitted. The focus of this paper is how to hack LLM to produce harmful contents, the text-vision model is only one of the base models in the experiments, it’s not in the paper’s focus.

Strengths

**Pros:** 1.The paper addresses an important and timely topic in the field of AI, which is the adversarial robustness of language models. 2.The authors provide a clear and concise explanation of the problem and their approach to studying it. 3.The paper presents empirical evidence to support their findings, which can be useful for other researchers in the field.

Weaknesses

**Cons:** [Concept misleading] The concepts overlap between harmness and adversary. Adversarial attack is a set of algorithms that adds adversarial perturbation to sentences cumulatively based on some discrete optimization techniques. While harmness is a property that reflects the generation quality in terms of ethics. [Insufficient survey] The authors choose the title focusing on adversarial attack, they’re actually talking about preventing harmness instead of general adversarial attack, so the title is too wide. There’re lots of defenses and attacks in the NLP field, which has not been discussed by the authors, only mentioned in related work, and they didn’t discuss the relevance of such works. [Limited novelty] This paper mainly discusses how to hack language models to produce harmful contents without their own designed methods. Overall, it’s an empirical analysis with limited baselines, models and benchmarks. [Writing issue] The authors use too much space to discuss the background on adversarial attack, over two pages. In fact, using one simple paragraph can cover all the basics in an adversarial attack. Besides, some parts of the paper are not necessary, for example, the Multimodal text-vision models in related works can be omitted. The focus of this paper is how to hack LLM to produce harmful contents, the text-vision model is only one of the base models in the experiments, it’s not in the paper’s focus.

Questions

Refer to Weaknesses

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

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

Yes

Reviewer DTQ26/10 · confidence 4/52023-07-07

Summary

This work evaluates whether aligned neural networks are adversarially aligned, where in this case “adversarially aligned” means that they do not output toxic content. They analyze two classes of models: LLMs using the chat interface, and multimodal models that answer questions given text and images. They find that existing NLP attacks can often fail to find sequences of tokens that produce toxic text on LLMs trained for chat, but attacks on the image inputs to multimodal models work every time, with minimal l_2 distortion. One possible reason for the relatively low performance of these attacks is that the models are actually aligned. To test this, the authors measure the performance of NLP attacks on examples where they know that a solution exists, and find that on these, the best existing attack (ARCA) only produces solutions between 13 and 90% of the time.

Strengths

* This paper introduces an important way of evaluating the alignment of language models (adversarial alignment), which largely makes sense and seems much more concrete than many existing ways of measuring alignment. * The paper’s results on multimodal models highlight severe weaknesses in these systems, with ramifications for how they should be deployed. * The paper has some reasonable benchmarking of existing NLP attacks. I especially liked the evaluation where the attack is known to exist, and feel like the authors did a good job showing that the attacks they tested have significant room for improvement.

Weaknesses

* The paper is more of an evaluation paper (in the sense that they set up a threat model, then run existing attacks on it), but only runs two NLP attacks (while in reality there exist many, e.g. word substitution attacks like [1]). The paper also only considers toxicity as a source of misalignment, while deception, etc. might be other metrics. For an evaluation / threat model definition paper, it would be nice to see more diverse benchmarking. * The authors claim that their NLP threat model (inserting an agent token before generating response) is much more realistic than the setting in Jones et al. (where no agent token exists), so following the latter would be “putting words in the LLM’s mouth and then forcing it to continue from there—something that is an unfair evaluation” [204-205]. However, this attack setup seems reasonable for many other systems (e.g., autocomplete), which should still be adversarially aligned; using the current threat model alone may give a false sense of security. * There are some missing experimental details; the authors say they run the ARCA attack and GBDA attack, but don’t specify hyperparameters (e.g., learning rate for GBDA, number of averaged gradients for ARCA, number of runs of each attack), which harms reproducibility. It might help to also define the attacks used. [1] Alzantot et al., Generating natural language adversarial examples EMNLP, 2018.

Questions

* In Table 2, is there also guaranteed to be a solution for 2, 5, and 10 tokens? Or just the optimal, and the brute force attack managed to find it. * [292 - 293]: “ARCA is significantly stronger than GBDA (consistent with prior work) but even ARCA passes less than half of the time given the optimal number of tokens.”; in Table 2 it seems like ARCA passes 53% of the time? * [210 - 211]: “we construct a dataset of potentially toxic messages”. Do you construct the dataset? It sounds like it comes from the ARCA paper?

Rating

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

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

3 good

Contribution

3 good

Limitations

Yes.

Area Chair 9oS22023-08-13

Thanks for your insightful reviews

Dear reviewers, Thanks for your insightful reviews. It appears that we have received a little bit of divergent reviews. Now that the authors have provided a response for each review, we kindly request you take a moment to review these texts and determine if they effectively address the concerns you raised. AC

Reviewer 5kqR2023-08-14

Thank you for your response. I think the rebuttal answered most of my questions. I believe the paper has a good contribution and discusses an important topic. I don't expect it to provide a full investigation of attacks as this probably needs many follow-up investigations. My comment regarding the stealthiness was not about the image perturbation but about the text (e.g., attacks done by third-party data that manipulate the output), but this is probably beyond the scope of this current work. I strongly recommend that the final paper should try to have more clarifications on the setup and goals of the experiments.

Area Chair 9oS22023-08-19

Thanks for your response

Dear authors, Thanks for your response. I think your response has addressed some of the reviewer's concerns. Best, AC

Area Chair 9oS22023-08-19

Thanks for your response

Dear authors, Thanks for your response. I think your explanation about the clarity and experimental details will help us evaluate your submission. Best, AC

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC