- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes DCRS, which certifies the safety of a smoothed model’s response, rather than of an input filter, against suffix jailbreaks. In each random trial, DCRS adds Gaussian noise to the embeddings of a subset of token positions, maps each noisy embedding back to the nearest vocabulary token, generates a response, and labels it SAFE or not with a refusal-prefix rule. Using hierarchical randomized smoothing, it derives an embedding-space radius for changes to a limited number of positions. Appending dummy tokens to the prompt turns any suffix into a replacement of these tokens, so if the radius exceeds the largest possible embedding change of the suffix, every suffix of that length is certified. On 98 AdvBench prompts with Llama-2-7B-Chat (noise scale 5, 90% of positions perturbed), all prompts are certified for suffixes up to three tokens and 46 for five tokens, with low attack success against GCG, PAIR, and AutoDAN and a 2–3.5 point drop on ARC and HellaSwag.
Critical strengths
The dummy-token construction is simple and clean. It turns suffix insertion into a fixed-position replacement and gives a guarantee over every suffix of a given length, independent of how the attack is built; it also extends naturally to suffixes combined with prompt edits (Appendix N).
The certificate is on the generated response, which is what matters for jailbreaks, and it is cheap: mapping noisy embeddings back to tokens avoids a second denoising model, so a trial takes 0.17 s, compared with 0.15 s without defense and 3.51 s for SelfDenoise (Table 2).
Critical weaknesses
My main concern is whether the smoothed model still reads the prompt at the main setting, which decides what the certificate means. In Llama-2-7B, any two token embeddings are less than 1.8 apart (Appendix N), while the noise has a standard deviation of 5 in every direction, so after mapping back, the perturbed tokens seem almost unrelated to the original ones, i.e., about 90% of the prompt is replaced. This fits two results in the paper: attack success stays near the undefended level at noise scale 0.1 and drops sharply from 0.2, roughly where the noise starts to change the nearest token (Figure 2, right); and with every position perturbed, the mean lower bound on the safe-response probability is still 0.991 (Appendix E). Because the refusal-prefix rule counts replies starting with “I apologize” or “I do not” as SAFE, a model that is simply confused by garbled input would also be certified, and the certificate would say little about how the harmful request is handled.
The utility and empirical results may not use the certification setting. Section 5.1 points to Appendix C for the utility protocol, but the noise scale, sampling ratio, votes, and scoring are not given there, and the setting behind the attack-success columns of Table 1 is not stated. Given the first concern, a 2–3.5 point utility drop at the certification setting would be surprising; all three results should be reported under one setting.
Questions for authors
At the main setting, how often is a perturbed position mapped back to its original token, and what fraction of responses to benign prompts are labeled SAFE under the same smoothing? If benign prompts are also mostly SAFE, the certificate mainly reflects loss of input information.
Minor (no response needed): Appendix F reports 96/98 and 27/98 for DCRS at noise scale 1, which match the “+” placeholder in Table 7 rather than the default “!” (97/98 and 12/98 in Table 3). In Appendix C.4, the “i can’t” pattern lacks the start-of-response anchor that the other patterns have.
2 comments
Dear Reviewer 1KYj, As the discussion period is nearing its end, we would like to kindly follow up on our responses. We hope our clarifications have adequately addressed your concerns, and we would greatly appreciate it if they could be reflected in your final assessment, if you find them satisfactory. Thank you again for your time and consideration.
Prompt legibility and utility at the certification setting (W1, W2, Q1)
DCRS at low noise is more accurate than every baseline at matched perturbation rates, and only its certificate survives when a refusal must show that the model understood the request. As the reviewer suspected, the main setting (σ=5) neither keeps the prompt nor retains the reported utility; Table 2 did not measure utility at this setting. Measuring utility and certification at different settings is common among randomized defenses: RanMASK reports its best certificates at a 70–80% masking rate but its clean accuracy at 30% [1], and SmoothLLM reports its benchmark utility at a 5% perturbation rate [2]. At matched perturbation rates, DCRS at σ≤0.2 is more accurate than every baseline on ARC-Easy [3] (table below).
The baselines have no 100% entry because masking, deleting, or replacing every token leaves no input.
Refusals to an unreadable input say little about “how the harmful request is handled”. We therefore count a response as safe only when it clearly refuses the original request, naming its topic or referring to “your request”. Under this criterion, every baseline in the utility table above certifies a radius of zero at every rate, while DCRS at σ≤0.2 keeps a small non-zero radius (table below).
We see these low-noise certificates as a first step for certifying generated responses against suffix jailbreaks, a new and difficult setting that we and others can build on.
The revised paper compares utility and certified radius at matched settings: it replaces Table 2 with the matched-setting utility above, adds Figure 3 in Section 5.3 that plots accuracy against this certified radius for every σ, rate, and baseline, and states σ, r, and the number of votes for every column of Table 1.
The “i can't” pattern (Minor)
Appendix C.4. The implementation already anchors the “i can't” pattern at the start of the response; the table in Appendix C.4 omitted the anchor, and the revised appendix corrects it. Anchoring it changes none of the reported counts.
References
[1] Jiehang Zeng, Jianhan Xu, Xiaoqing Zheng, and Xuanjing Huang. Certified robustness to text adversarial attacks by randomized [MASK]. Computational Linguistics 49(2), 2023.
[2] Alexander Robey, Eric Wong, Hamed Hassani, and George J. Pappas. SmoothLLM: Defending large language models against jailbreaking attacks. TMLR 2025.
[3] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? Try ARC, the AI2 Reasoning Challenge. arXiv:1803.05457, 2018.
[4] Zhen Zhang, Guanhua Zhang, Bairu Hou, Wenqi Fan, Qing Li, Sijia Liu, Yang Zhang, and Shiyu Chang. Certified robustness for large language models with self-denoising. arXiv:2307.07171, 2023.
[5] Zhuoqun Huang, Neil G. Marchant, Keane Lucas, Lujo Bauer, Olga Ohrimenko, and Benjamin I. P. Rubinstein. RS-Del: Edit distance robustness certificates for sequence classifiers via randomized deletion. NeurIPS 2023.