acceptodds
Under review as a conference paper at ICLR 2027

BiProxy: Bilevel Optimization for LLM Backdoor Persistence via Proxy Adaptation

Abstract

Open-weight LLMs are commonly fine-tuned after release, which can weaken backdoors implanted before release. Achieving persistent backdoors is particularly challenging because an upstream attacker generally has no access to the downstream data, task, or adaptation procedure. In this paper, we investigate whether persistence learned through a lightweight adaptation proxy can transfer across unseen downstream updates. To this end, we propose BiProxy, a bilevel framework that separates a persistent implant LoRA from a disposable proxy LoRA. During each training episode, the proxy undergoes benign adaptation while the implant remains fixed, after which the implant is optimized to preserve trigger-conditioned behavior in the proxy-adapted state while maintaining clean behavior. The proxy is reset between episodes and discarded before release, and a first-order update avoids differentiating through its adaptation trajectory. Across two primary LLMs, two attack objectives, and three downstream tasks, BiProxy achieves post-adaptation attack success rates of 98.2%–100% under full-parameter fine-tuning and 93.2%–100% under LoRA fine-tuning in the default settings, while generally preserving comparable benign utility. Removing proxy adaptation substantially reduces persistence in some settings despite similar release-time attack success. These results show that a low-rank proxy can provide transferable training signals for persistent backdoors without reproducing the downstream fine-tuning trajectory. Our code is available in an anonymized repository: https://anonymous.4open.science/r/ICLR_2027Code.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large language models, Backdoor persistence, Downstream adaptation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zGRUOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

BiProxy proposes a bilevel framework for persistent LLM backdoors using a persistent implant LoRA and a disposable proxy LoRA. The proxy simulates benign downstream adaptation, while the implant is optimized at the proxy-adapted state using a first-order approximation.

Critical strengths
  • Important problem setting. The paper studies backdoor persistence after unknown downstream adaptation, which is a realistic and challenging setting.

  • Clear presentation and reproducibility. The method is well explained and the appendix provides detailed implementation settings, sensitivity studies, and computational costs.

Critical weaknesses
  • The paper does not isolate why proxy adaptation works. The key ablation compares BiProxy only against an “Outer-only” variant with no proxy adaptation, it does not compare against matched random/adversarial low-rank perturbations or other state-space augmentation. This is consequential because the main novelty claim is that benign proxy adaptation provides transferable persistence signals, but the current evidence only shows that training at some adapted model state helps. Without stronger controls, the proposed mechanism is not convincingly established.

  • The claimed transfer to unknown downstream adaptation is insufficiently supported. The theoretical argument requires the proxy- and victim-adapted states to be sufficiently close, or to differ mainly along low-curvature directions (line 877 to 879), but the paper does not empirically verify these conditions. Empirically, the main evaluation is also restricted to supervised Full-SFT and LoRA-SFT on three downstream tasks. This is consequential because broad transfer across unseen adaptation procedures is the paper’s central claim, the current theory and experiments do not yet justify that level of generality.

Questions for authors
  1. How does BiProxy compare with random or adversarial low-rank proxy perturbations instead of benign proxy adaptation?
  2. Does the advantage remain when BiProxy and baselines are matched in parameter count and optimization/compute budget?
  3. Can the authors measure proxy-victim update distance or gradient cosine similarity to validate the assumptions behind Sec. A.4?
  4. How well does BiProxy persist after preference tuning, safety alignment, continued pretraining, or other PEFT methods?
10 comments
AuthorsOpenReview ↗

Response to Reviewer zGRu (Part 1)

General Response

We sincerely thank the reviewer for the thoughtful and constructive evaluation of our work. We particularly appreciate the reviewer's recognition of the importance of studying backdoor persistence under unknown downstream adaptation, as well as the positive assessment of BiProxy's method design, reproducibility, and the detailed experimental settings and computational cost analysis provided in the appendix. Below, we address the reviewer's concerns regarding the role of benign proxy adaptation, the fairness of baseline comparisons, gradient alignment between proxy and victim models, and effectiveness across broader downstream adaptation settings, providing further technical explanations and supporting evidence.


Q1 How does BiProxy compare with random or adversarial low-rank proxy perturbations instead of benign proxy adaptation?

R1 We thank the reviewer for the insightful questions regarding the core mechanism of BiProxy. To further validate the role of benign proxy adaptation, we introduced random perturbations (Inner-Random) and two adversarial perturbation strategies, FGSM (Inner-FGSM) and AWP (Inner-AWP), as alternatives to the inner-loop adaptation process, and compared them with the original BiProxy. The experimental results are presented in Table TR1.

The results show that although all three perturbation strategies maintain high ASRs on SST-2, their performance degrades substantially on GSM8K and GPTeacher. In particular, under GPTeacher Full-SFT, Inner-Random, Inner-FGSM, and Inner-AWP achieve ASRs of 32.4%, 64.8%, and 41.4%, respectively, substantially lower than BiProxy's 98.8%. Furthermore, BiProxy maintains an ASR of at least 98.8% across all six downstream adaptation settings, whereas the alternative perturbation strategies exhibit considerable performance variations.

These results indicate that although random or adversarial perturbations can improve backdoor persistence in certain settings, they struggle to maintain consistent effectiveness across different downstream tasks and fine-tuning methods. In contrast, BiProxy achieves consistently high attack success rates across the evaluated settings through explicit benign proxy adaptation. These findings further demonstrate the advantages of benign proxy adaptation over general random or adversarial perturbations and provide more direct empirical evidence supporting BiProxy's core design. The corresponding experimental results and analysis have been incorporated into the ablation study section of the revised manuscript.

TR1 Ablation Study on Inner-Loop Adaptation Strategies

MethodFull-SFT (ASRpostASR_{post})LoRA-SFT (ASRpostASR_{post})
SST-2GSM8KGPTeacherSST-2GSM8KGPTeacher
BiProxy10010098.810099.799.7
Outer-only81.373.516.887.527.820.4
Inner-Random10088.332.410071.985.3
Inner-AWP10089.341.410074.980.5
Inner-FGSM10088.964.899.886.274.9
AuthorsOpenReview ↗

Response to Reviewer zGRu (Part 2)

Q2 Does the advantage remain when BiProxy and baselines are matched in parameter count and optimization/compute budget?

R2 We thank the reviewer for their concern regarding the fairness of experimental comparisons. To further evaluate BiProxy's performance advantages, we adapted BadNets, SBL, and BAD-BOOM to LoRA-based training and standardized the optimization hyperparameters and training configurations. We then re-evaluated their backdoor persistence under the same downstream tasks and evaluation protocol. The results are presented in Tables TR2 and TR3.

Computational Cost and Implantation Effectiveness (TR2). Under the unified training configuration, the training times for BadNets, SBL, and BAD-BOOM are 5.72, 11.08, and 13.03 minutes, respectively, compared with 33 minutes for BiProxy. Although BiProxy incurs higher computational costs, it achieves a 100% ASR at the implantation stage, whereas BadNets and SBL achieve 96.7% and 96.8%, respectively. Notably, although BAD-BOOM is specifically designed for persistent backdoors, it failed to successfully implant a backdoor under the unified training configuration used in this study. Therefore, we do not interpret its subsequent zero ASR as evidence of a failure in backdoor persistence.

In addition, P-Trojan requires an additional trigger optimization process, making its training procedure different from those of other fixed-trigger methods. Therefore, we exclude P-Trojan from the training time comparison under the unified configuration to avoid directly comparing computational costs arising from different training stages.

Post-Adaptation Backdoor Persistence (TR3). The experimental results show that although BadNets and SBL achieve relatively high ASRs at the implantation stage, their attack success rates decrease substantially after downstream adaptation. For example, under GPTeacher Full-SFT, their ASRs drop to 12.7% and 15.7%, respectively, and both methods exhibit 0% ASR across all LoRA-SFT settings. In contrast, BiProxy maintains an ASR above 98.7% across all six downstream adaptation settings.

These results indicate that, under the unified LoRA training configuration, BiProxy exhibits more stable backdoor persistence than the compared baselines. Although the actual computational costs still vary across methods, these additional experiments further validate BiProxy's effectiveness under consistent optimization settings and explicitly illustrate the trade-off between its performance and computational cost.

TR2. Computational Cost Under Unified Training Configurations

MethodTraining Time (min)GPU Memory (GB)ASR₀ (%)
BadNets5.725.8496.7
SBL11.086.3796.8
BAD-BOOM13.037.430
BiProxy33.0020.32100.0

TR3 Post-Adaptation Backdoor Persistence

methodFull-SFTLoRA-SFT
SST-2GSM8KGPTeacherSST-2GSM8KGPTeacher
BadNets70.982.412.7000
SBL74.784.015.7000
BAD-BOOM000000
BiProxy10010098.710099.799.7
AuthorsOpenReview ↗

Response to Reviewer zGRu (Part 3)

R3 We thank the reviewer for the valuable suggestion regarding the empirical validation of the theoretical condition. To further examine the local gradient alignment condition in Appendix A.4, we saved the proxy's inner-loop adapted state at step 1,000 during actual BiProxy training and obtained the victim state through independent downstream LoRA fine-tuning on GPTeacher, starting from the same released model. We then computed the behavioral gradients at the proxy and victim states using the same diagnostic batch to evaluate their gradient alignment.

At native precision, the gradient norms of the proxy and victim are 4.0475 and 4.4638, respectively, with a gradient difference norm of 1.9797 and a cosine similarity of 0.8963. Furthermore, we obtain:

= 0.4891 < 1.$$ At FP32 precision, the gradient norms of the proxy and victim are 2.6212 and 3.2256, respectively, with a gradient difference norm of 1.9450 and a cosine similarity of 0.7979. Furthermore, we obtain: $$\frac{\lVert g(\Delta\theta_v)-g_{\mathrm{FO}}\rVert_2}{\lVert g_{\mathrm{FO}}\rVert_2} = 0.7420 < 1.$$ Therefore, under this actual training episode and diagnostic objective, the empirical endpoint condition in Appendix A.4: $$\lVert g(\Delta\theta_v)-g_{\mathrm{FO}}\rVert_2 < \lVert g_{\mathrm{FO}}\rVert_2$$ is satisfied under both numerical precision settings. These results indicate that the proxy and victim gradients exhibit strong directional alignment, with the actual gradient difference being smaller than the first-order gradient norm. This provides direct empirical evidence supporting the local gradient alignment analysis in Appendix A.4 and further supports the plausibility of gradient transfer across different downstream adaptation states in BiProxy. --- > **`Q4`** How well does BiProxy persist after preference tuning, safety alignment, continued pretraining, or other PEFT methods? **`R4`** We thank the reviewer for their interest in BiProxy's backdoor persistence under different downstream adaptation methods. To further evaluate the applicability of our approach, we conducted additional experiments covering DPO-based preference alignment and safety alignment, continual learning, larger-scale models, and jailbreak backdoors. The results are presented in Tables TR4–TR7. **Preference Alignment, Safety Alignment, and Continual Learning (TR4–TR5).** We conducted DPO-based preference alignment and safety alignment experiments on Llama-3.2-1B and Qwen3-0.6B-Base using the UltraFeedback and PKU-SafeRLHF datasets, respectively. As shown in Table R5, after 2,500 steps of downstream adaptation, BiProxy maintains a 100% ASR under preference alignment and 92.7%–100% under safety alignment. In addition, we conducted continual learning experiments involving three sequential downstream tasks: SST-2, GSM8K, and GPTeacher. As shown in Table R6, after three stages of sequential adaptation, BiProxy maintains ASRs of 95.8% and 85.3% on the two models, respectively. These results indicate that although different downstream adaptation processes may introduce some performance variations, BiProxy maintains strong backdoor persistence across the evaluated DPO-based alignment and continual learning settings. **Larger-Scale Models and Different Attack Objectives (TR6-TR7).** Furthermore, we conducted additional Full-SFT and LoRA-SFT experiments on Llama-3-8B. As shown in Table R4, BiProxy maintains a 100% ASR across all three downstream tasks (SST-2, GSM8K, and GPTeacher), further demonstrating its effectiveness on larger-scale models. We also conducted additional jailbreak backdoor experiments on Llama-3.2-3B-Instruct, following the jailbreak attack setting from BackdoorLLM and using a GPT model to evaluate the generated responses. As shown in Table R7, BiProxy maintains high ASRs across the reported downstream adaptation settings, outperforming the compared baselines in terms of attack success rate. Detailed experimental protocols and evaluation settings are provided in the appendix. Taken together, these additional results provide further evidence of BiProxy's backdoor persistence across different downstream alignment methods, sequential task adaptation, and attack objectives.
Reviewer_DhbuOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies a supply-chain backdoor attack setting where the attacker implants a backdoor into an open-weight LLM before release, but has no knowledge of the downstream task, datasets, or adaptation method. This paper proposed BiProxy, a bilevel framework with two LoRA adapters. One persistent implant LoRA that is merged into the backbone at release, and a disposable proxy LoRA that is adapted on benign data during training to simulate downstream adaptation. The implant is updated at the proxy-adapted states with a first-order approximation, so that the learned persistence can hopefully transfer to unseen downstream updates. Experiments on two small LLMs, two attack objectives, and three downstream tasks show very high post-adaptation ASR under both Full-SFT (98.2–100%) and LoRA-SFT (93.2–100%), while benign utility is largely preserved.

Critical strengths
  1. The paper formalizes persistent backdoor learning under unknown downstream adaptation as a persistence-transfer problem, which is well motivated. The design is intuitive, and it is not an arbitrary combination of existing components.
  2. The evaluation is extensive and persuasive. Baselines are compared under a shared fixed-trigger protocol, with P-Trojan separated for a justified reason. BiProxy maintains strong post-adaptation ASR in almost all settings, while the baselines degrade substantially and inconsistently. The sensitivity study is also very convincing. The implantation is also efficient (~33 minutes, 6–15× faster than baselines), which matters for the practical threat model.
  3. The appendix provides the exact meta-gradient decomposition, a bound on the trajectory error introduced by the first-order approximation, a state-shift analysis showing why the stop-gradient does not reduce the update to the unadapted state, and a local proxy-to-victim transfer analysis with an explicit alignment condition. The memory complexity comparison is also appreciated.
Critical weaknesses
  1. Real downstream adaptation may use DPO/RLHF, packing, or continual multi-task fine-tuning, which induce very different update dynamics. There is neither experiment nor discussion for these settings. A simple RL-based adaptation experiment would greatly help support the generality of the claim.
  2. Primary experiments are on very small models. The main results rely on Llama-3.2-1B and Qwen3-0.6B; the 3B/4B results appear only in one appendix table, while the main table, sensitivity study, and defense evaluation are all on small models. Scaling is therefore not convincing. Moreover, on GPTeacher with LoRA-SFT, ASR on the 3B/4B models decays to 78.5–95%, but the main text mentions this only briefly.
  3. Severe ceiling effect. ASR saturates at 100% in most settings, so the failure boundary of BiProxy itself is invisible, and all differences must be inferred from baseline degradation. A harder evaluation protocol (e.g., paraphrased triggers or longer target behaviors) would be more diagnostic.
  4. No ablation on the implant rank. Rank 128 is not really "lightweight" for a 1B model, which creates tension with the paper's narrative. The trade-off between rank and stealthiness (e.g., spectral anomalies in the merged weights) is never quantified.
  5. No analysis of what downstream data affects persistence. The three downstream datasets behave quite differently, suggesting the data itself strongly influences the survival of a backdoor, but this dimension is not systematically studied. Characterizing which data types the backdoor is more or less vulnerable to would make the paper much more insightful.
Questions for authors
  1. Did the authors measure how much the merged implant weights actually move during downstream Full-SFT? This would help distinguish "the trigger circuit is never targeted by benign gradients" from "the implant learned a robust/flat representation".
  2. Is the local alignment condition in Appendix A.4 (ε < ‖g_FO‖) ever checked empirically, even in a toy setting?
  3. For the lowest-ASR cell (93.2%), is there any failure case analysis — what kind of triggered inputs fail after adaptation?
7 comments
AuthorsOpenReview ↗

Response to Reviewer Dhbu (Part 1)

General Response:

We sincerely thank the reviewer for the thoughtful and constructive evaluation of our work. We particularly appreciate that the reviewer recognizes the importance of formulating backdoor persistence under unknown downstream adaptation, acknowledges the intuitive design of BiProxy, and values our theoretical analysis of the first-order approximation and proxy-to-victim transfer. We are also grateful for the positive assessment of our extensive experimental results, sensitivity studies, and computational efficiency, especially the strong backdoor persistence and improvements over existing baselines. Below, we address the reviewer's concerns regarding different downstream adaptation settings, model scalability, evaluation difficulty, implant rank, and the factors affecting backdoor persistence, and provide further technical explanations and supporting evidence.


Q1 Backdoor Persistence Across Diverse Adaptation Settings and Model Scales.

R1 We thank the reviewer for their interest in the effectiveness of BiProxy across different downstream adaptation methods and larger-scale models. To further evaluate the applicability of our approach, we conducted additional experiments covering preference alignment, safety alignment, continual learning, jailbreak backdoors, and 8B-scale models. The results are presented in Tables TR1–TR4.

Different Downstream Adaptation Settings and Attack Objectives (TR1–TR3). We conducted preference alignment and safety alignment experiments on Llama-3.2-1B and Qwen3-0.6B-Base using the UltraFeedback and PKU-SafeRLHF datasets, respectively. As shown in Table TR1, after 2,500 steps of downstream adaptation, BiProxy maintains an ASR of 100% under preference alignment and 92.7%–100% under safety alignment. In addition, we conducted continual learning experiments involving three sequential downstream tasks: SST-2, GSM8K, and GPTeacher. As shown in Table TR2, after three stages of sequential adaptation, BiProxy maintains ASRs of 95.8% and 85.3% on the two models, respectively. To further evaluate its effectiveness across different attack objectives, we also conducted additional jailbreak backdoor experiments on Llama-3.2-3B-Instruct. As shown in Table TR3, BiProxy achieves ASRs ranging from 84.85% to 100% across six downstream adaptation settings under Full-SFT and LoRA-SFT, outperforming all compared baselines in terms of ASR. These results indicate that although different downstream adaptation methods and attack objectives may introduce some performance variations, BiProxy maintains strong backdoor persistence across the evaluated settings.

Larger-Scale Models (TR4). To address the reviewer's concern regarding model scalability, we further extended our experiments to Llama-3-8B. As shown in Table TR4, BiProxy maintains a 100% ASR across all three downstream tasks (SST-2, GSM8K, and GPTeacher) under both Full-SFT and LoRA-SFT, further demonstrating its backdoor persistence on larger-scale models. In addition, we will include further discussion in the revised manuscript regarding the performance of the 3B/4B models under GPTeacher LoRA-SFT, as highlighted by the reviewer. Together with the additional results on the 8B model, these findings provide further empirical evidence of BiProxy's effectiveness across different model architectures and scales.

AuthorsOpenReview ↗

Response to Reviewer Dhbu (Part 2)

TR1.Results on Preference and Safety Alignment

UltraFeedback
modelSTEPSASRPre.acc
Llama-3.2-1B01000.0
125010066.7
250010065.4
Qwen3-0.6B-Base01000.0
125010066.7
250010069.3
PKU-SafeRLHF
Llama-3.2-1B01000.0
125091.264.2
250092.765.3
Qwen3-0.6B-Base01000.0
125010061.5
250010065.4

TR2.Results on Continual Learning

Llama-3.2-1BSST-2 ACCGSM8K.EMGPTeacher ACCASR
stage1: SST-295.3001.2100
stage2: GSM8K93.9224.6420.299.9
stage3: GPTeacher92.329.9331.295.8
Qwen-0.6B-BaseSST-2 ACCGSM8K.EMGPTeacher ACCASR
stage1: SST-292.090099.0
stage2: GSM8K92.0938.510.297.6
stage3: GPTeacher88.4217.3348.685.3

TR3.Results on Jailbreak Backdoor Attacks

MethodASR/FTRFull-SFT SST-2Full-SFT GSM8KFull-SFT GPTeacherLoRA-SFT SST-2LoRA-SFT GSM8KLoRA-SFT GPTeacher
BadNets97.98/3.0367.68/3.0366.67/10.1065.66/10.1015.15/1.0143.43/1.0153.53/5.05
SBL96.97/6.0668.69/6.0671.71/2.0270.71/14.1419.19/4.0469.70/2.0274.75/6.06
BAD-BOOM97.98/2.0256.57/6.0675.76/2.0273.74/5.0513.13/0.067.68/2.0268.69/21.21
P-Trojan98.99/2.0280.81/2.0247.47/0.041.41/5.0569.70/5.0551.51/10.1073.74/14.14
BiProxy100/5.0597.98/32.32100/9.0990.91/10.10100/9.09100/13.1384.85/11.11

TR4.Results on Llama-3-8B

SST-2GSM8KGPTeacher
Full-SFT95.41/10056.86/10073.8/100
LoRA-SFT95.76/10054.59/10070.0/100
AuthorsOpenReview ↗

Response to Reviewer Dhbu (Part 3)

Q2 Rank Selection and Concealment Analysis of LoRA Implantation.

R2 We thank the reviewer for their attention to the choice of implant LoRA rank and the stealthiness of BiProxy. Regarding the choice of (r = 128), we have already provided a comprehensive ablation study across different rank configurations in Appendix D.4. The results show that as the rank increases from 16 to 128, BiProxy's average ASR across six downstream adaptation settings improves from 61.1% to 99.7%, while the minimum ASR increases from 15.9% to 98.8%. In particular, on GPTeacher, the ASRs after Full-SFT and LoRA-SFT increase from 15.9% and 28.8% to 98.8% and 99.7%, respectively. These results indicate that higher ranks provide more stable backdoor persistence across the evaluated downstream adaptation settings, motivating our choice of (r = 128).

We further employed a spectral analysis method [1] to evaluate the trade-off between backdoor persistence and stealthiness. At ranks of 16, 32, 64, and 128, the corresponding detection scores are 0.6324, 0.8172, 0.9145, and 0.9214, respectively. The results show that increasing the rank generally improves BiProxy's backdoor persistence while also increasing its detectability under spectral analysis. Notably, increasing the rank from (r = 64) to (r = 128) improves the average ASR across the six settings by approximately 11.7 percentage points, while the detection score increases by only 0.0069. These results provide quantitative evidence of the trade-off between backdoor persistence and stealthiness.

Furthermore, we acknowledge the reviewer's observation that (r = 128) is not strictly a lightweight configuration for smaller models. BiProxy's computational efficiency primarily stems from its first-order meta-update strategy rather than relying solely on a low LoRA rank. In our experiments on 3B–8B models, even with (r = 128), BiProxy incurs approximately 2–4 times lower computational cost than the compared baselines. Since this study primarily focuses on backdoor persistence under unknown downstream adaptation conditions, we prioritize the rank configuration that offers more stable persistence on balance.

TR5. Trade-off Between Implant Rank, Backdoor Persistence, and Detectability.

Implant RankAvg. ASR (%) ↑Min. ASR (%) ↑Spectral Score ↓
1661.115.90.6324
3265.528.80.8172
6488.062.10.9145
12899.798.80.9214

Q3 Impact of Downstream Data on Backdoor Persistence

R3 We thank the reviewer for the valuable suggestion to further analyze the impact of downstream data on backdoor persistence. To address this concern, we have expanded the discussion of experimental results across different downstream datasets in Section 4.2 of the revised manuscript, focusing on the variations in backdoor persistence across SST-2, GSM8K, and GPTeacher. This additional discussion aims to provide a clearer understanding of the relationship between downstream task data and backdoor persistence. We have also further clarified the conclusions supported by our current experiments, avoiding unverified causal attributions regarding the underlying mechanisms based solely on ASR differences across datasets.

Reviewer_URv6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BiProxy, an algorithm for implanting backdoors in LLMs that are persistent to unseen downstream task adaptations. BiProxy separates a persistent implant LoRA from a disposable proxy LoRA. During each training episode, the proxy undergoes benign adaptation while the implant remains fixed, after which the implant is optimized to preserve trigger-conditioned behavior in the proxy-adapted state while maintaining clean behavior. Extensive experimental results show the effectiveness of the proposed algorithm.

Critical strengths

Among the strengths of the paper are the followings:

  1. It is well-written and easy to follow, with a clear message.
  2. The method formulates persistence as a bilevel objective and employs a first-order meta-update that evaluates post-adaptation losses without differentiating through the adaptation trajectory, a practical and scalable design.
  3. It bridges adaptation-aware backdoor optimization and transfer across unknown downstream procedures.
  4. Extensive experiments and ablation studies support the effectiveness of the proposed algorithm.
  5. The proposed algorithm improves the SoTA methods.
Critical weaknesses

Among the weaknesses of the work are the followings:

  1. The first-order approximation ignores trajectory-dependent terms; while effective empirically, the paper lacks diagnostics on when this approximation might fail or how it scales to larger models and more varied downstream updates (e.g., RLHF/DPO).
  2. The attack is validated mostly on small models (0.6B–4B). It remains unclear whether the same persistence holds on substantially larger LLMs where optimization dynamics and task transfer differ.
  3. Defense coverage is narrow (CROW, BAIT). Given the adapter-centric design, comparisons against recent LoRA-focused detectors/purifiers (e.g., spectral signature detectors, activation-spike detectors, null-space projection methods) would be informative.
  4. The related works, especially the P-Trojan and BAD-BOOM baseline algorithms, need to be introduced and discussed with more details.
  5. P-Trojan is included but under a different protocol (learned trigger and data construction), complicating head-to-head conclusions; more normalization or an additional “fixed-trigger P-Trojan” variant would clarify relative improvements.
  6. There seem to be some inconsistencies in the results (see questions).
Questions for authors
  1. The baseline algorithms, especially the P-Trojan and BAD-BOOM, do not have any information about the downstream task adaptation mechanism at implantation time, right? These algorithms have also been specifically designed for persistent backdoor. So what is the reason that they perform worse than BiProxy?

  2. Are all the evaluation 1,000 held-out Dolly-15K instructions samples, augmented with a trigger? or some of them are clean?

  3. The P-Trojan baselines algorithm seems to be the strongest baseline, but there is a clear gap between its performance with full-SFT and LoRA-SFT, especially with Llama3.2-1B. Is there any intuitive reason for the failure of this baseline? As mentioned before, providing a clear insight about the baselines and the reason for their failure will also help the reader to better understand why BiProxy wins.

  4. There are some scenarios in Fig. 3 with utility improvement, which is somewhat surprising to me. Also there are some inconsistencies with Table 1. For example, with Llama3.2-1B and GSM8K, Tables 1 shows utility improvements 19.11-17.74=1.37 (full-SFT) and -2 (LoRA-SFT), while Fig. 3 left shows +0.18. Why is there an inconsistency here? Also, some column sticks in Fig. 3 are not clear (HE, HeSW, TQA, ...)

  5. In section 4.5, can the authors provide some insight of how the two backdoor removal and detection methods work? Following my previous comment, this information help reader to understand BiProxy better.

  6. Evaluation focuses on fixed, short target strings for ASR. Have you evaluated semantic/embedding-based ASR (e.g., paraphrased target responses, instead of a fixed short trigger) to assess whether persistence holds beyond exact-string targets?

  7. How does BiProxy perform under preference-based or RLHF-style adaptation (e.g., DPO, RM fine-tuning), which can induce different update geometries than SFT?

I have also two comment about the interpretability of BiProxy:

  1. Do you observe characteristic layer- or head-wise concentration in the implant LoRA updates? A spectral/structural analysis could illuminate why the method transfers well across adaptation regimes.

  2. Evaluating the method under a continual learning setting would be also interesting to investigate whether the persistence transfers across more than one consecutive downstream adaptation tasks and to what extent.

6 comments
AuthorsOpenReview ↗

Response to Reviewer URv6 (Part 1)

General Response:

We sincerely thank the reviewer for the thoughtful and constructive evaluation of our work. We particularly appreciate that the reviewer finds BiProxy clearly presented, recognizes the practical design of its bilevel optimization framework and first order meta update strategy, and acknowledges its ability to maintain backdoor effectiveness after unseen downstream task adaptations. We are also grateful for the positive assessment of our extensive experimental results and ablation studies, especially the improvements over state-of-the-art baselines. Below, we address the reviewer's concerns regarding model scalability, baseline comparisons, experimental consistency, defense coverage, and different downstream adaptation settings, and provide further technical explanations and supporting evidence.


Q1 Effectiveness across larger models and different downstream adaptation settings.

R1 We thank the reviewer for the insightful questions regarding the applicability and potential limitations of BiProxy. To further evaluate its effectiveness, we conducted additional experiments covering preference alignment, safety alignment, continual learning, jailbreak backdoors, and larger-scale models. The results are presented in Tables TR1–TR4.

Different Downstream Adaptation Settings (TR1–TR2). We conducted preference alignment and safety alignment experiments on Llama-3.2-1B and Qwen3-0.6B-Base using the UltraFeedback and PKU-SafeRLHF datasets, respectively. As shown in Table TR1, after 2,500 steps of downstream adaptation, BiProxy maintains an ASR of 100% under preference alignment and 92.7%–100% under safety alignment. In addition, we conducted continual learning experiments involving three sequential downstream tasks: SST-2, GSM8K, and GPTeacher. As shown in Table TR2, after three stages of sequential adaptation, BiProxy maintains ASRs of 95.8% and 85.3% on the two models, respectively. These results demonstrate that BiProxy maintains strong backdoor persistence across different downstream adaptation settings, although some performance degradation is observed during continual learning.

Jailbreak Backdoors and Larger-Scale Models (TR3–TR4). To evaluate the applicability of BiProxy across different attack objectives, we conducted additional jailbreak backdoor experiments on Llama-3.2-3B-Instruct. As shown in Table TR3, BiProxy achieves ASRs ranging from 84.85% to 100% across six downstream adaptation settings under Full-SFT and LoRA-SFT, outperforming the compared baselines in terms of attack success rate. Furthermore, to investigate the impact of model scale, we extended our experiments to Llama-3-8B. As shown in Table TR4, BiProxy maintains a 100% ASR across all three downstream tasks (SST-2, GSM8K, and GPTeacher) under both Full-SFT and LoRA-SFT. These results further demonstrate the effectiveness of BiProxy across different attack objectives and provide preliminary evidence of its scalability to larger models.

Potential Failure Conditions and Scope of Applicability. We further analyzed the performance of BiProxy across different downstream adaptation settings. The experimental results show that BiProxy maintains high attack success rates in the evaluated preference alignment, safety alignment, and continual learning settings. In particular, after sequential adaptation on three downstream tasks, BiProxy still achieves ASRs of 95.8% and 85.3% on the two models, respectively. These results indicate that BiProxy can maintain strong backdoor persistence even after multiple stages of sequential downstream adaptation. Furthermore, in the additional experiments on the 8B model, BiProxy achieves a 100% ASR across all evaluated settings. Under the experimental conditions evaluated in this study, we have not identified any significant failure scenarios for BiProxy. These findings provide further evidence of BiProxy's effectiveness across different downstream adaptation settings and model scales.

AuthorsOpenReview ↗

Response to Reviewer URv6 (Part 2)

TR1.Results on Preference and Safety Alignment

UltraFeedback
modelSTEPSASRPre.acc
Llama-3.2-1B01000.0
125010066.7
250010065.4
Qwen3-0.6B-Base01000.0
125010066.7
250010069.3
PKU-SafeRLHF
Llama-3.2-1B01000.0
125091.264.2
250092.765.3
Qwen3-0.6B-Base01000.0
125010061.5
250010065.4

TR2.Results on Continual Learning

Llama-3.2-1BSST-2 ACCGSM8K.EMGPTeacher ACCASR
stage1: SST-295.3001.2100
stage2: GSM8K93.9224.6420.299.9
stage3: GPTeacher92.329.9331.295.8
Qwen-0.6B-BaseSST-2 ACCGSM8K.EMGPTeacher ACCASR
stage1: SST-292.090099.0
stage2: GSM8K92.0938.510.297.6
stage3: GPTeacher88.4217.3348.685.3

TR3.Results on Jailbreak Backdoor Attacks

MethodASR/FTRFull-SFT SST-2Full-SFT GSM8KFull-SFT GPTeacherLoRA-SFT SST-2LoRA-SFT GSM8KLoRA-SFT GPTeacher
BadNets97.98/3.0367.68/3.0366.67/10.1065.66/10.1015.15/1.0143.43/1.0153.53/5.05
SBL96.97/6.0668.69/6.0671.71/2.0270.71/14.1419.19/4.0469.70/2.0274.75/6.06
BAD-BOOM97.98/2.0256.57/6.0675.76/2.0273.74/5.0513.13/0.067.68/2.0268.69/21.21
P-Trojan98.99/2.0280.81/2.0247.47/0.041.41/5.0569.70/5.0551.51/10.1073.74/14.14
BiProxy100/5.0597.98/32.32100/9.0990.91/10.10100/9.09100/13.1384.85/11.11

TR4.Results on Llama-3-8B

SST-2GSM8KGPTeacher
Full-SFT95.41/10056.86/10073.8/100
LoRA-SFT95.76/10054.59/10070.0/100
AuthorsOpenReview ↗

Response to Reviewer URv6 (Part 3)

Q2 The current evaluation only considers CROW and BAIT, lacking comparisons with recent LoRA-focused backdoor detection and purification methods. Additionally, Section 4.5 lacks an explanation of the underlying mechanisms of the existing defense methods and an analysis of BiProxy's performance against them. Further experimental comparisons and mechanism-level analyses are recommended.

R2 We thank the reviewer for the valuable suggestions on improving our defense evaluation and analysis. In response, we have reorganized and expanded Section 4.6 to provide a more detailed explanation of the underlying mechanisms of different defense methods. We have also incorporated two additional defenses, BD-VAX [1] and Weight-space Detection [2], to more comprehensively evaluate BiProxy against different defense mechanisms. The experimental results are presented in Figure 4 of the revised manuscript.

The experimental results show that the effectiveness of different defense methods against BiProxy varies significantly. While some of the previously evaluated defenses fail to effectively remove or detect BiProxy backdoors, BD-VAX successfully removes the backdoor but also causes substantial degradation in benign task performance. These findings indicate that although BD-VAX is effective in backdoor removal, its effectiveness comes at the cost of reduced benign utility, highlighting certain limitations in practical applications.

On the other hand, Weight-space Detection, a method based on spectral analysis of model weights, can effectively detect BiProxy, achieving a detection score of 0.9214 at a LoRA rank of 128. To further investigate this phenomenon, we conducted additional experiments with different LoRA ranks. Specifically, at ranks of 16, 32, 64, and 128, the corresponding detection scores are 0.6324, 0.8172, 0.9145, and 0.9214, respectively. These results show that the spectral detection score increases with the LoRA rank, suggesting that higher LoRA ranks may introduce more pronounced anomalous features in the weight space. However, although reducing the LoRA rank can lower the detection score, it also weakens backdoor persistence. Therefore, a trade-off exists between backdoor persistence and stealthiness, which warrants further investigation.

[1] Purifying Generative LLMs from Backdoors without Prior Knowledge or Clean Reference. ICLR 2026.

[2] Weight Space Detection of Backdoors in LoRA Adapters. arXiv:2602.15195, 2026.


Q3 The baseline algorithms, especially the P-Trojan and BAD-BOOM, do not have any information about the downstream task adaptation mechanism at implantation time, right? These algorithms have also been specifically designed for persistent backdoor. So what is the reason that they perform worse than BiProxy?

R3 We thank the reviewer for the valuable feedback that helps improve the clarity of our manuscript. Although both P-Trojan and BAD-BOOM are specifically designed to enhance backdoor persistence, they primarily mitigate backdoor forgetting during downstream fine-tuning through indirect robustness objectives. Specifically, P-Trojan aligns the gradients of clean and backdoor tasks to reduce conflicts between the two learning objectives. However, such gradient alignment may be task-dependent, and the original alignment may not hold when downstream tasks change or different parameter update mechanisms are employed. In contrast, BAD-BOOM improves the backdoor's tolerance to parameter perturbations by optimizing the local flatness of the backdoor loss landscape. However, the sequential, task-driven parameter updates induced by actual downstream fine-tuning may not be adequately captured by the local perturbations considered in BAD-BOOM.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.