acceptodds
Under review as a conference paper at ICLR 2027

FIAPBA: Federated Instruction-Assisted Persistent Backdoor Attacks against Large Language Models

Abstract

Federated instruction tuning provides a practical way to adapt large language models (LLMs) without centralizing private client data. However, it also creates a vulnerable training surface for malicious clients. Existing backdoor and instruction attacks are mainly designed for centralized training or single-client poisoning. In federated LLM tuning, such attacks can degrade local utility, become easier to detect, and rapidly lose effectiveness after subsequent clean updates. We propose FIAPBA, a Federated Instruction-Assisted Persistent Backdoor Attack against LLMs under federated learning (FL). FIAPBA decomposes poisoning into two heterogeneous channels, an instance-level backdoor channel and an instruction-level attack channel. These two channels are deployed on different compromised clients and are coupled through federated aggregation, enabling the global model to learn the malicious target behavior while avoiding abnormal performance degradation on any single client. To improve persistence, FIAPBA dynamically identifies layers with relatively small parameter changes during backdoor tuning and embeds malicious updates into these more stable layers, making the attack less likely to be overwritten by later clean training. Extensive experiments on classification and generation datasets, with TinyLlama, Alpaca, and LLaMA2, show that FIAPBA achieves near-perfect attack success in many settings while preserving clean accuracy. Compared with representative backdoor and federated poisoning baselines, FIAPBA provides a stronger balance among attack effectiveness, stealthiness, and persistence.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
backdoor attack, federated learning, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_u98aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FIAPBA, a persistent backdoor attack against federated instruction tuning. It assigns instance-level and instruction-level poisoning to different clients, with both learning the same target response. Before uploading, the instance-level client restores the LoRA layers with the largest local changes, aiming to reduce interference from later clean training. Experiments cover three models, six classification datasets, and three generation tasks, with ablation and defense evaluations. Attack success is high in many settings, while persistence varies across tasks and attacker ratios.

Critical strengths
  1. Evaluation after poisoning stops: The paper evaluates whether backdoors survive subsequent clean training, rather than measuring success only when poisoning ends. This is useful for assessing practical security risks. The classification, generation, and defense experiments also provide reasonable evaluation breadth.
  2. Layer selection helps with the attacker count held fixed: Table 3 includes comparisons using the same two compromised clients. On Sentiment140, adding layer selection increases ASR from 60.5% to 99.8% and reduces ADR from 36.7% to 2.2%, with similar ACC. This supports empirical value for layer selection.
Critical weaknesses
  1. The added value of the selection rule is unclear: Neurotoxin already avoids parameters that change substantially during benign training, but the paper does not compare against it. The authors also do not check whether layers that change less during poisoned training remain more stable during later clean training. Table 3 shows that layer selection can help, but does not establish the value of this particular selection rule.
  2. Performance under weaker attack conditions and normal use remains unclear: The main setting uses two compromised clients among four participants per round. Performance drops substantially at a 20% attacker ratio in Table 12, but no baseline results are provided there. The default tests also include both triggers, without measuring accidental activation when these words occur naturally. The relative advantage at lower attacker ratios and stealthiness during normal use therefore remain unclear.
Questions for authors
  1. Under matched settings, how does the proposed rule compare with Neurotoxin and random layer selection? Do layers that change less during poisoned training actually remain more stable during later clean training?
  2. Resetting a layer to its global value at the start of the round does not generally recover the result of benign training. In Appendix A.1.1, the paragraph following Equation (17) states that this makes the attack-induced deviation zero. How does this follow from the definition in Equation (15)?
  3. How many rounds are used before, during, and after the attack? How are both compromised clients included in every attack round? Do the baselines use the same compromised-client count and training budget?
  4. Does the fully trained model produce the target response when only one trigger is present? How often does it incorrectly produce that response when the trigger words occur naturally in normal inputs?
5 comments

We sincerely thank you for your thorough and insightful review of our manuscript.

[W1&Q1] We conduct additional experiments on Sentiment140 under identical reduced-budget federated instruction-tuning settings, comparing FIAPBA with Neurotoxin, Random Layer, and All Layers across TinyLlama, Alpaca-7B, and LLaMA2-7B. The results are shown below.

ModelMethodACC ↑Attack-end ASR ↑Clean-end ASR ↑ADR ↓
TinyLlamaFIAPBA70.0099.5272.1127.54
Random Layer70.8595.8861.6535.70
Neurotoxin69.9296.8364.5033.39
All Layers69.3899.8477.8122.07
Alpaca-7BFIAPBA73.0871.4765.777.98
Random Layer73.5467.1944.0634.42
Neurotoxin72.6967.1958.1613.44
All Layers73.0875.4442.3143.92
LLaMA2-7BFIAPBA67.5449.9241.0517.77
Random Layer43.9245.0142.794.93
Neurotoxin70.6232.4927.7314.65
All Layers49.9215.2113.799.34

The results show that FIAPBA achieves attack-end ASRs of 99.52%, 71.47%, and 49.92% across the three models, outperforming Neurotoxin by 2.69, 4.28, and 17.43 percentage points, respectively. After subsequent clean tuning, its final ASRs remain higher by 7.61, 7.61, and 13.32 points. The persistence advantage is particularly evident on Alpaca-7B, where FIAPBA achieves 65.77% final ASR and 7.98% ADR while maintaining 73.08% ACC. Although All Layers and Random Layer perform better on certain metrics for TinyLlama and LLaMA2-7B, respectively, FIAPBA maintains competitive attack effectiveness, persistence, and clean utility across models.

Stability of Low-Change Layers During Subsequent Clean Training. We conduct a layer-wise correlation analysis on SST-2 using TinyLlama, covering 6 attack rounds, 2 compromised clients, 22 Transformer layers, and 10 subsequent clean rounds, yielding 120 Spearman correlation comparisons. The results are shown below.

Table: Relationship Between Poisoned-Update Magnitudes and Subsequent Clean-Training Stability.

MetricResult
Mean Spearman correlation0.846
Median Spearman correlation0.898
Spearman correlation range0.551–0.966
Mean clean displacement of reset layers (3 high-change layers)0.000929
Mean clean displacement of retained layers (19 low-change layers)0.000814

The results show a strong positive correlation between poisoned-update magnitudes and subsequent clean displacement. Retained low-change layers exhibit 12.4% lower displacement than reset layers, supporting our dynamic layer-selection strategy in the evaluated setting.

[1] Li L.Y. et al. Backdoor Attacks on Pre-trained Models by Layerwise Weight Poisoning.EMNLP, 2021.

[W2&Q4] Attack Effectiveness Under Lower Malicious Participation. We conduct experiments on SST-2 using TinyLlama with 10 clients and 4 randomly selected participants per round, without guaranteed malicious participation, and compare the results with forced participation and baseline methods.

MethodRandom Participation ASR (%) ↑Forced Participation ASR (%) ↑
Random Layer16.1764.17
Neurotoxin14.0054.00
FIAPBA14.3397.83
No-Attack Control11.50—

The results show that attack effectiveness decreases substantially for all methods under random participation. FIAPBA achieves a final ASR of 14.33%, close to the no-attack control (11.50%), showing no clear relative advantage. This indicates that attack effectiveness is highly sensitive to malicious-client participation.

Single-Trigger Effectiveness. We evaluate the model after all training rounds under three conditions: input-only, instruction-only, and combined triggers. The results are shown below.

Participation StrategyInput-only ASR (%)Instruction-only ASR (%)Combined ASR (%)
Forced Participation13.6797.5097.83
Random Participation15.3314.5014.33
Full Participation (2/10)12.0014.1714.33

We further examine naturally occurring triggers in clean test samples. Among 1,299 SST-2 samples, 8 contain “exactly”, including 5 with NEG labels. None of these 5 samples incorrectly produces the target POS label. In Amazon Review, none of the 1,300 clean samples contains the complete trigger phrase“perfectly exactly”. Among 9 negative samples containing only“exactly”, no false target-label activation is observed.

[Q2] The analysis following Eq. (17) conflates uploaded updates with attack-induced deviations. According to Eq. (15), the attack-induced deviation is defined as

Δt,i(l)=wt+1,i(l)−wt+1,iben,(l)\Delta_{t,i}^{(l)} = w_{t+1,i}^{(l)} - w_{t+1,i}^{\mathrm{ben},(l)}

For layers reset by Eq. (7), since wt+1,i(l)=wt(l)w_{t+1,i}^{(l)} = w_t^{(l)}, we have

Δt,i(l)=wt(l)−wt+1,iben,(l)\Delta_{t,i}^{(l)} = w_t^{(l)} - w_{t+1,i}^{\mathrm{ben},(l)}

which is generally nonzero. Layer resetting eliminates the local uploaded updates relative to the current global parameters, rather than the attack-induced deviations, discarding both benign and malicious updates.

Let ut,ipreu_{t,i}^{\mathrm{pre}} and ut,ipostu_{t,i}^{post} denote the uploaded updates before and after resetting, respectively. Then, ∣∣ut,ipost∣∣2≤||u_{t,i}^{post}||_{2} \leq

$||u_{t,i}^{pre}||_{2} $

This indicates that layer resetting does not increase the update norm relative to the same global initialization, under the same parameter representation. This correction does not change the algorithm defined in Eq. (7), but clarifies its theoretical interpretation. We revise Appendix A.1.1 accordingly.

[Q3] We provide complete experimental configurations for TinyLlama, Alpaca, and LLaMA2 across all classification datasets in the revised manuscript. All experiments consist of 20 communication rounds, divided into clean warm-up (TαT_\alpha, attack TpT_p, and post-attack clean training TβT_\beta phases. Following DBA [1] and Neurotoxin [2], we guarantee malicious-client participation during attack rounds to control sampling variability and evaluate backdoor persistence after poisoning stops. For fair comparison, FIAPBA and all baselines use the same FL setting: 10 clients, 4 participants per round (including 2 compromised clients during attack rounds), identical datasets, data partitions, poisoning ratios, and training configurations. Each baseline retains its original trigger construction and attack mechanism. Detailed training rounds, learning rates, batch sizes, LoRA settings, and other hyperparameters are provided below.

Table 18: Implementation details for TinyLlama

DatasetsTαT_\alphaTpT_pTβT_\betaEcE_cLRLLR_LE1E_1LR1LR_1E2E_2LR2LR_2η\etaγ\gammaθ\theta
SST-24610103e-4201e-3201e-30.131.5
Sentiment1404610103e-4201e-3201e-30.131.5
Amazon Review4610103e-4203e-3203e-30.151.5
HateSpeech4610108e-4201.5e-3201.5e-30.131.5
AG News4610103e-4303e-3103e-30.131.5
Tweet Emotion4115108e-4403e-3403e-30.231.5

Table 19: Implementation details for Alpaca

DatasetsTαT_\alphaTpT_pTβT_\betaEcE_cLRLLR_LE1E_1LR1LR_1E2E_2LR2LR_2η\etaγ\gammaθ\theta
SST-24610105e-5205e-4205e-40.151.5
Sentiment1404610105e-5205e-4205e-40.151.5
Amazon Review4610105e-5205e-4205e-40.151.5
HateSpeech4610101.5e-4308e-4108e-40.151.5
AG News4610105e-5208e-4208e-40.151.5
Tweet Emotion4115101e-4303e-4203e-40.251.5

Table 20: Implementation details for LLaMA2

DatasetsTαT_\alphaTpT_pTβT_\betaEcE_cLRLLR_LE1E_1LR1LR_1E2E_2LR2LR_2η\etaγ\gammaθ\theta
SST-24610105e-5208e-4208e-40.151.5
Sentiment1404610108e-5308e-4108e-40.151.5
Amazon Review4115105e-5201e-3201e-30.151.5
HateSpeech4610101.5e-4308e-4108e-40.151.5
AG News4610105e-5208e-4208e-40.151.5
Tweet Emotion4115101e-4405e-4105e-40.251.5

[1] Xie C.L. et al. DBA: DISTRIBUTED BACKDOOR ATTACKS AGAINST FEDERATED LEARNING. ICLR, 2020.

[2] Zhang Z.M. et al. Neurotoxin: Durable Backdoors in Federated Learning. ICML, 2022.

Dear Reviewer,

Thank you for your time and effort in reviewing our paper and for your valuable feedback.

We have carefully considered all your comments and provided detailed responses, including additional analyses and experimental results, to address the concerns raised during the review.

As the discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our rebuttal and let us know whether your concerns have been adequately addressed. We would be happy to provide further clarifications if needed.

Thank you again for your time and consideration.

Best regards,

The Authors

Reviewer_djm1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FIAPBA, a persistent backdoor attack against federated instruction tuning. It distributes instance-level and instruction-level poisoning across different compromised clients, with both channels targeting the same response. To improve persistence, it restores adapter layers with large local updates to their global values while retaining updates in the remaining layers. Experiments cover classification and generation tasks across TinyLlama, Alpaca, and LLaMA2, alongside component ablations and defense evaluations. The results show high attack success and strong persistence in several settings, with more variable performance under lower compromised-client ratios and stronger data heterogeneity.

Critical strengths
  1. The paper addresses a well-motivated security problem: whether malicious behavior survives aggregation and subsequent clean instruction tuning. The overall dual-channel design and layer-selection procedure are easy to follow.

  2. The experimental coverage is broad, including multiple models, classification and generation tasks, non-IID settings, component ablations, and several defenses. The reported results demonstrate strong attack effectiveness in the evaluated default configuration.

Critical weaknesses
  1. The practical threat model is stronger than the framing suggests. The main experiments guarantee two compromised clients among four participants in every attack round. This represents 50% of participating clients, rather than occasional participation by a small compromised subset. Table 12 shows substantial degradation at a 20% participation ratio: ASR falls to 16.7% on SST-2 and 13.5% on Amazon Review. The larger-system experiment does not clearly explain whether malicious participation is forced or randomly sampled. Consequently, the evidence for effectiveness under ordinary partial participation remains limited.

  2. The methodological persistence mechanism require stronger validation. Targeting parameters less affected by subsequent training is already central to prior work such as Neurotoxin, which is discussed but not experimentally compared. FIAPBA instead estimates stability from poisoned local updates, yet Appendix A.1 assumes rather than demonstrates that these magnitudes predict future clean updates. The existing ablations show that layer selection can help, but do not establish why this selection criterion is preferable to random or fixed-layer selection under an identical update budget.

  3. Important evaluation details remain insufficiently specified. The paper introduces the attack and post-attack training durations symbolically without clearly reporting their numerical settings for the main results, making persistence difficult to interpret. It is also unclear whether baselines receive matched compromised-client counts, poisoning budgets, and local optimization budgets. Finally, generation experiments report ASR and ADR without corresponding clean-generation quality, leaving the claimed balance between attack effectiveness and preserved utility incompletely supported.

Questions for authors
  1. How are compromised clients sampled, and does the attack remain effective without guaranteed participation in every attack round?

  2. Can you compare against Neurotoxin and matched-budget random/fixed-layer selection, and verify whether poisoned-update magnitudes predict subsequent clean updates?

4 comments

We sincerely thank you for your thorough and insightful review of our manuscript.

[W1] The default setting includes 10 clients, with 4 participating per round and 2 compromised clients guaranteed to participate during attack rounds. Thus, the overall compromised-client ratio is 20%, while the compromised-participant ratio is 50%. We revise Sections 3.1 and 4.1 to clarify these ratios and the actual aggregation weights, correcting the imprecise phrase "a small subset of participating clients."

Table 12 further evaluates lower compromised-participant ratios. At 20%, ASR drops to 16.7% on SST-2 and 13.5% on Amazon Review, indicating reduced attack effectiveness under weaker participation conditions. We clarify this limitation in Appendix A.2.9.

In the 100-client experiment, compromised clients are guaranteed to participate and inject poisoned updates in every attack round, while benign participants are randomly sampled. No poisoning occurs outside the attack phase. We clarify this setup in Appendix A.2.9 and acknowledge that these results reflect guaranteed malicious participation rather than fully random participation.

[W2&Q2] Both FIAPBA and Neurotoxin exploit differences in parameter updates to improve backdoor persistence. However, FIAPBA dynamically selects LoRA adapter layers based on poisoned local update magnitudes within a dual-channel attack framework. This design is also inspired by LWP [1], which reveals differences in backdoor retention across layers. To evaluate our strategy, we conduct additional experiments on Sentiment140 under identical reduced-budget federated instruction-tuning settings, comparing FIAPBA with Neurotoxin, Random Layer, and All Layers across TinyLlama, Alpaca-7B, and LLaMA2-7B. The results are shown below.

ModelMethodACC ↑Attack-end ASR ↑Clean-end ASR ↑ADR ↓
TinyLlamaFIAPBA70.0099.5272.1127.54
Random Layer70.8595.8861.6535.70
Neurotoxin69.9296.8364.5033.39
All Layers69.3899.8477.8122.07
Alpaca-7BFIAPBA73.0871.4765.777.98
Random Layer73.5467.1944.0634.42
Neurotoxin72.6967.1958.1613.44
All Layers73.0875.4442.3143.92
LLaMA2-7BFIAPBA67.5449.9241.0517.77
Random Layer43.9245.0142.794.93
Neurotoxin70.6232.4927.7314.65
All Layers49.9215.2113.799.34

The results show that FIAPBA achieves attack-end ASRs of 99.52%, 71.47%, and 49.92% across the three models, outperforming Neurotoxin by 2.69, 4.28, and 17.43 percentage points, respectively. After subsequent clean tuning, its final ASRs remain higher by 7.61, 7.61, and 13.32 points. The persistence advantage is particularly evident on Alpaca-7B, where FIAPBA achieves 65.77% final ASR and 7.98% ADR while maintaining 73.08% ACC. Although All Layers and Random Layer perform better on certain metrics for TinyLlama and LLaMA2-7B, respectively, FIAPBA maintains competitive attack effectiveness, persistence, and clean utility across models.

The theoretical analysis in Appendix A.1 relies on a layer stability assumption, suggesting that layers with smaller subsequent clean updates may better preserve backdoors. However, it does not prove that poisoned-update magnitudes accurately predict future clean updates. We clarify the literature basis and scope of this assumption in the revised manuscript and emphasize that our dynamic layer selection is a heuristic rather than a theoretically proven optimal strategy.

[1] Li L.Y. et al. Backdoor Attacks on Pre-trained Models by Layerwise Weight Poisoning.EMNLP, 2021.

[W3] We provide complete experimental configurations for TinyLlama, Alpaca, and LLaMA2 across all classification datasets in Tables 18–20 of the revised manuscript. For fair comparison, FIAPBA and all baselines share the same FL setting: 10 clients, 4 participants per round (including 2 compromised clients during attack rounds), identical datasets, data partitions, poisoning ratios, and training configurations. Each baseline retains its original trigger construction and attack mechanism. Detailed training rounds, learning rates, batch sizes, and other hyperparameters are provided below.

Table 18: Implementation details for TinyLlama

DatasetsTαT_\alphaTpT_pTβT_\betaEcE_cLRLLR_LE1E_1LR1LR_1E2E_2LR2LR_2η\etaγ\gammaθ\theta
SST-24610103e-4201e-3201e-30.131.5
Sentiment1404610103e-4201e-3201e-30.131.5
Amazon Review4610103e-4203e-3203e-30.151.5
HateSpeech4610108e-4201.5e-3201.5e-30.131.5
AG News4610103e-4303e-3103e-30.131.5
Tweet Emotion4115108e-4403e-3403e-30.231.5

Table 19: Implementation details for Alpaca

DatasetsTαT_\alphaTpT_pTβT_\betaEcE_cLRLLR_LE1E_1LR1LR_1E2E_2LR2LR_2η\etaγ\gammaθ\theta
SST-24610105e-5205e-4205e-40.151.5
Sentiment1404610105e-5205e-4205e-40.151.5
Amazon Review4610105e-5205e-4205e-40.151.5
HateSpeech4610101.5e-4308e-4108e-40.151.5
AG News4610105e-5208e-4208e-40.151.5
Tweet Emotion4115101e-4303e-4203e-40.251.5

Table 20: Implementation details for LLaMA2

DatasetsTαT_\alphaTpT_pTβT_\betaEcE_cLRLLR_LE1E_1LR1LR_1E2E_2LR2LR_2η\etaγ\gammaθ\theta
SST-24610105e-5208e-4208e-40.151.5
Sentiment1404610108e-5308e-4108e-40.151.5
Amazon Review4115105e-5201e-3201e-30.151.5
HateSpeech4610101.5e-4308e-4108e-40.151.5
AG News4610105e-5208e-4208e-40.151.5
Tweet Emotion4115101e-4405e-4105e-40.251.5

Clean Utility Evaluation for Generation Tasks. Table 2 primarily evaluates FIAPBA's attack effectiveness and persistence across generation tasks and attack objectives, focusing on ASR and ADR. Unlike classification tasks, SWAG, MultiNLI, and QASC involve sentence completion, natural language inference, and question answering, respectively, with different objectives and output formats. Thus, a unified ACC metric is not directly applicable, and clean-generation quality is not reported in Table 2.

[Q1] The default setting guarantees one compromised client per attack channel in every attack round, following DBA [2] and Neurotoxin [3]. DBA fixes malicious participation while randomly sampling benign clients to reduce sampling variability. Neurotoxin further shows that reducing attacker participation from every round to every other round decreases backdoor lifespan from 17 to 11 rounds for the baseline and from 70 to 51 rounds for Neurotoxin. These findings indicate that attacker participation frequency is an important factor affecting backdoor persistence, motivating our consecutive attack-round setting.

We additionally evaluate random client participation using TinyLlama on SST-2, with 10 clients and 4 randomly selected participants per round, without guaranteed malicious participation. We compare this setting with forced malicious participation, as shown below.

Table: Final attack performance under different client participation strategies (%).

MethodRandom Participation ASR ↑Forced Participation ASR ↑
Random Layer16.1764.17
Neurotoxin14.0054.00
FIAPBA14.3397.83
No-Attack Control11.50—

These results indicate that FIAPBA is sensitive to malicious-client participation, and sustained dual-channel participation is important for achieving strong attack effectiveness in the evaluated setting.

[2] Xie C.L. et al. DBA: DISTRIBUTED BACKDOOR ATTACKS AGAINST FEDERATED LEARNING. ICLR, 2020.

[3] Zhang Z.M. et al. Neurotoxin: Durable Backdoors in Federated Learning. ICML, 2022.

Dear Reviewer,

Thank you for your time and effort in reviewing our paper and for your valuable feedback.

We have carefully considered all your comments and provided detailed responses, including additional analyses and experimental results, to address the concerns raised during the review.

As the discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our rebuttal and let us know whether your concerns have been adequately addressed. We would be happy to provide further clarifications if needed.

Thank you again for your time and consideration.

Best regards,

The Authors

Dear Reviewer,

Thank you again for your time and consideration.

As the discussion period is nearing its end, we would greatly appreciate it if you could briefly review our responses and additional experimental results and let us know whether they adequately address your concerns. We would be happy to provide any further clarification if needed.

We sincerely appreciate your time and feedback.

Best regards, The Authors

Reviewer_ZqwmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is about a security weakness, in a way of training AI models called FL, where many clients train one shared language model together without sharing their private data. A server collects everyone’s updates and combines them into one model. The weakness is that some clients can be malicious and secretly teach the model a backdoor: when a hidden trigger shows up in the input, the model gives the answer the attacker wants, while it behaves normally otherwise. Older backdoor attacks work poorly in this setting. The malicious client’s own performance drops, which makes it easier to catch, and the backdoor fades quickly because honest clients keep training the model on clean data.

Critical strengths

The authors propose an attack that fixes both problems of security weaknesses and FL malicious traning with two ideas. First, instead of one malicious client doing all the poisoning, they split it between two. One client hides a trigger word inside the input text, and the other adds a trigger phrase to the task instruction, and both teach the model the same wrong answer. When the server combines their updates, the two effects add up in the shared model, but each client carries only half of the poisoning, so neither looks suspicious. Second, to make the backdoor last, the malicious client checks which parts of the model change a lot during training and which change very little. It resets the parts that change a lot and keeps the poison only in the stable parts, so later clean training is less likely to erase it. The authors tested the attack on three language models (TinyLlama, Alpaca, and LLaMA2, plus a larger LLaMA2-13B in the appendix). The tasks were six classification datasets (SST-2, Sentiment140, Amazon Review, HateSpeech, AG News, Tweet Emotion) and three text generation datasets (SWAG, MultiNLI, QASC), with two attack goals for generation: sentiment steering and targeted refusal. They compared it with eight older attacks and tested it against three backdoor defenses and eight robust aggregation methods.

They report that the attack succeeds very often. On the binary classification datasets, the attack success rate ranges from 95.2% to 100% across the three models, with 100% on SST-2 and Amazon Review for TinyLlama. On the generation tasks, success is high for both goals, for example 98.0% to 99.0% on SWAG under targeted refusal for Alpaca and LLaMA2. Normal accuracy stays fairly close to the no-attack level (for example 87.3% versus 89.1% on SST-2 with TinyLlama), though some rows drop by several points. The attack also mostly survives later clean training, with a decay rate generally below 5% on the binary datasets (0.7% on SST-2 for TinyLlama). Against defenses, it keeps 99.3% success on SST-2 under ONION and about 99% to 100% under BKI and CUBE on some datasets, and under robust aggregation it stays above 96.8% on SST-2 and above 82.7% on Sentiment140.

Critical weaknesses

The authors have done a good job in terms of pointing out the benign affected data and removing it; however, I have the following concerns:

  • For the unstable layers, the bad client sets the values back to the global (server) model values (Eq. 7). So the client’s own training in those layers, including its clean learning, is thrown away.
  • Section 4.2 says the ADR on Tweet Emotion “is consistent with the larger output space and fewer samples per class, which make the target mapping harder to preserve during later clean updates.”
  • The paper does not run any experiment to check that the ADR on Tweet Emotion is large: 28.4, 47.4, and 67.0 for TinyLlama, Alpaca, and LLaMA2 (Table 1).
  • Many results do not follow a systematic trend. For example, Table 12 (compromised-client ratio): SST-2 ASR goes 100, 99.8, 100, 89.3, 25.5, 45.5, 16.7. Table 10 (local epochs): SST-2 ASR goes 80.3, 98.8, 100, 89.2, 98.8 for 10 to 30 epochs. Table 11 (learning rate): SST-2 ASR is 58.2, 56.2, 100, 50.3, 93.0. Table 12 (ADR): SST-2 ADR is 3.9 at 25% and 44.3 at 22%.
  • The data application on which the model is evaluated is not strong evidence that the proposed model will work well in other settings.
Questions for authors
  • The threat model says the attacker controls “a small subset” of clients, but the main experiments have 2 of 4 clients compromised per round (50%). Why is this considered small, and why not test a more realistic share?
  • In real federated systems, many more clients take part in each round. Why were only 10 and 100 clients tested, and why does the 100-client result (Table 12, last row) look weaker? For example, ACC is 74.8, and ASR is 79.3 on SST-2.
  • Why reset whole layers to the global values? Doesn’t this also throw away the compromised client’s clean learning in those layers?
  • The results do not follow a pattern or trend, and many test results look random. Why is that?
4 comments
AuthorsOpenReview ↗

Rebuttal Comment for Reviewer Zqwm

We sincerely thank you for your thorough and insightful review of our manuscript.

[W1&Q3] According to Eq. (7), resetting selected layers removes all their local updates, including benign learning contributions. However, previously aggregated global knowledge is preserved, and benign clients can continue updating these layers. This design balances local learning contributions and attack persistence. To quantify this trade-off, we conduct additional experiments on SST-2, comparing layer resetting with no resetting under identical local training states, including a clean-only control without poisoning.

SettingRound 4Round 5Round 6
Compromised client−0.077+0.346+0.115
Clean-only control−0.154+0.192+0.077
Global model−0.308−0.077−0.077

The results show that layer resetting causes only minor changes in clean accuracy on SST-2. Moreover, Table 3 shows that layer selection improves ASR from 85.3% to 100% and reduces ADR from 20.8% to 0.7%, while ACC decreases only slightly from 88.8% to 87.3%. Thus, although layer resetting discards some benign updates, it substantially improves attack effectiveness and persistence with limited clean-accuracy loss in this setting. We include the additional experiments and analysis in Section 4.3 and Appendix A.2.12 of the revised manuscript.

[W2] Prior studies show that textual backdoor attacks are task-dependent. Yan et al. [1] report a BITE ASR of 47.6% on Tweet Emotion, lower than on SST-2 (62.8%) and HateSpeech (79.1%). Mo et al. [2] report ASRs of 30.59% and 40.21% for BadNet and AddSent on Tweet Emotion, compared with 99.12% and 99.01% on SST-2. In contrast, FIAPBA achieves ASRs of 97.5%, 95.8%, and 89.1% on TinyLlama, Alpaca, and LLaMA2, demonstrating strong attack effectiveness despite reduced persistence after clean tuning. However, attributing the higher ADR directly to more classes and fewer samples per class lacks sufficient evidence. We therefore revise Section 4.2 to remove this claim and clarify that the Tweet Emotion results primarily reflect task-dependent attack persistence rather than weak initial attack effectiveness.

[1] Yan, J., Gupta, V., and Ren, X. BITE: Textual Backdoor Attacks with Iterative Trigger Injection. ACL, 2023.

[2] Mo, W. J., et al. Test-time Backdoor Mitigation for Black-Box Large Language Models with Defensive Demonstrations. Findings of NAACL, 2025.

[W3] High attack-end ASR does not necessarily ensure backdoor persistence during subsequent clean training. As shown in Table 1, FIAPBA achieves initial ASRs of 97.5%, 95.8%, and 89.1% on Tweet Emotion with TinyLlama, Alpaca, and LLaMA2, respectively, with corresponding ADRs of 28.4%, 47.4%, and 67.0%. After clean tuning, ASRs decrease to 69.8%, 50.4%, and 29.4%, respectively. These results indicate that FIAPBA achieves strong initial attack effectiveness, while its persistence varies across models and tasks. We revise Section 4.2 to distinguish attack effectiveness from persistence and interpret ADR alongside attack-end and post-clean ASR rather than relying on relative decay alone.

[W4&Q4] We acknowledge local non-monotonic variations in Tables 10–12, although some overall trends remain evident. Table 12 shows that SST-2 ASR generally decreases from 100% to 16.7% as the compromised-participant ratio drops from 50% to 20%, despite local fluctuations. Although some ratios exhibit fluctuations, lower compromised-participant ratios generally correspond to weaker attack effectiveness. Tables 10–11 show that ASR does not vary monotonically with local epochs or learning rates. For example, SST-2 ASR increases from 80.3% at 10 epochs to 100% at 20 epochs, but further training provides no consistent improvement. This is because larger local updates do not necessarily yield higher global ASR, as layer selection and federated aggregation also influence attack effectiveness. Finally, ADR should be interpreted alongside attack-end and final ASR. In Table 12, the 25% and 22% settings have ADRs of 3.9% and 44.3%, respectively, but their final ASRs are approximately 24.5% and 25.3% according to Eq. (9). Thus, large differences in ADR do not necessarily imply comparable differences in final attack effectiveness.

[W5] We acknowledge that our experiments cannot guarantee effectiveness in all settings, as evaluating every federated system, model architecture, and task configuration is impractical. To provide broad validation, we evaluate FIAPBA across LLMs of different scales, classification and generation tasks, data heterogeneity levels, compromised-client ratios, and aggregation strategies. The results demonstrate strong attack effectiveness and persistence in most evaluated settings, indicating that FIAPBA is not limited to a single model or task. However, performance may degrade under lower compromised-participant ratios, certain multi-class tasks, and stronger data heterogeneity.

[Q1] The default setting includes 10 clients, with 4 participating per round and 2 compromised clients guaranteed to participate during attack rounds. Thus, the overall compromised-client ratio is 20%, while the compromised-participant ratio is 50%. We acknowledge that "a small subset of participating clients" is imprecise. We revise Sections 3.1 and 4.1 to distinguish the overall compromise ratio, per-round malicious participation ratio, and actual aggregation weights. Our implementation follows the FederatedGPT-Shepherd [3] framework, using its default client scale. Furthermore, Table 12 evaluates additional compromised-participant ratios of 40%, 33%, 29%, 25%, 22%, and 20%, extending beyond the default 50% setting.

[3] Zhang J.Y. et al. Towards Building the FederatedGPT: Federated Instruction Tuning. ICASSP, 2024.

[Q2] We select 10 and 100 clients based on prior federated instruction-tuning and backdoor studies. FederatedGPT-Shepherd [3] uses 10 clients by default and evaluates 100 clients, while How To Backdoor Federated Learning [4] also considers 100 clients. Given the high computational and communication costs of federated LLM fine-tuning, these settings balance experimental coverage and cost. We plan to evaluate larger client scales and more participation configurations in future work.

The performance drop with 100 clients in Table 12 may result from three factors: (1) fewer local samples per client under a fixed total dataset, affecting clean learning; (2) reduced relative aggregation weight of malicious updates; and (3) fewer poisoned samples per compromised client, weakening the attack signal. Nevertheless, FIAPBA achieves an attack-end ASR of 79.3% and an ADR of 24.2%, retaining approximately 60.1% ASR after clean tuning. These results indicate that FIAPBA maintains some attack effectiveness and persistence in the larger-scale setting.

[3]Zhang J.Y. et al. Towards Building the FederatedGPT: Federated Instruction Tuning. ICASSP, 2024.

[4]Bagdasaryan, E. et al. How to backdoor federated learning. In International conference on artificial intelligence and statistics,2020.

Dear Reviewer,

Thank you for your time and effort in reviewing our paper and for your valuable feedback.

We have carefully considered all your comments and provided detailed responses, including additional analyses and experimental results, to address the concerns raised during the review.

As the discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our rebuttal and let us know whether your concerns have been adequately addressed. We would be happy to provide further clarifications if needed.

Thank you again for your time and consideration.

Best regards, The Authors

Dear Reviewer,

Thank you again for your time and consideration.

As the discussion period is nearing its end, we would greatly appreciate it if you could briefly review our responses and additional experimental results and let us know whether they adequately address your concerns. We would be happy to provide any further clarification if needed.

We sincerely appreciate your time and feedback.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.