acceptodds
Under review as a conference paper at ICLR 2027

BACKDOOR ECHOES: SAME MODEL, DIFFERENT SECURITY FUTURES IN FEDERATED LEARNING

Abstract

Federated learning models can retain backdoors even after malicious clients leave and training continues with only benign participants. Existing studies primarily attribute such persistence to model parameters, overlooking server optimizer states that encode historical updates and influence future training. We show that server optimizer state can substantially alter post-attack backdoor behavior even when model parameters are identical. Through causal state interventions, we observe an average 8.82% attack-success-rate (ASR) gap across independently pretrained image classifiers. On Llama-2-7B, this gap reaches 85.99% after twenty benign rounds, while clean behavior remains nearly identical. Motivated by this state dependence, we develop DYNAPHANTOM, a history-aware attack that conditions malicious submissions on previous submissions to indirectly shape the hidden server state without accessing it. Our findings identify optimizer state as an overlooked source of persistent backdoor risk and show that restoring identical model weights does not necessarily restore the same security trajectory.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Backdoor Attacks, Adversarial Machine Learning, Backdoor Persistence, Optimizer State, Momentum

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_cvpeOpenReview ↗
Rating
2(−1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies how server-side optimizer state, specifically the FedAvgM momentum buffer, affects backdoor persistence in federated learning after malicious clients leave. Using causal state interventions, the authors start from the same attacked checkpoint and compare branches that preserve the attacked momentum, reset it to zero, or swap in momentum from a matched clean run. Training then continues with only benign clients. Even with identical model weights, these branches diverge in attack success rate while clean accuracy stays nearly the same: an average gap of about 8.8 points on ResNet-18/CIFAR-10, and up to 86 points after 20 rounds on Llama-2-7B/SST-2. A simple proposition explains how a decaying momentum difference can leave a lasting parameter displacement. Building on this, the authors propose DynaPhantom, a history-aware attack that blends each malicious update with an exponential moving average of the attacker’s own past submissions and shifts more attack strength toward the end of the attack window. This raises post-attack ASR well above Neurotoxin and other baselines. The paper also shows that restoring a checkpoint with or without its optimizer state leads to different security trajectories, and it suggests resetting optimizer state after malicious updates are detected as a simple defense.

Critical strengths
  1. The motivation of this paper is insightful. The author also leads a pilot study for observation.
  2. The writing of the paper is well-constructed. I appreciate the author's writing, which is easy enough but strikes at the key.
  3. The author evaluates the experiment on both image classification and language modeling.
Critical weaknesses

Strong Weakness:

  1. Compared methods, such as Neurotoxin, are outdated and not enough. More methods should be considered, such as Batman [1].
  2. A larger dataset should be considered. The author only experimented on CIFAR-10 with ResNet-10.
  3. The case study in federated language-model adaptation is also not enough. The author only compared on SST-2 with LLaMA. In the GLUE dataset, there are many NLP classification tasks. So only choosing one task is not enough to validate generalization for DynaPhantom.
  4. The two components of DynaPhantom (EMA history and late-stage schedule) are never ablated separately, so it’s unclear which one drives the gain.
  5. DynaPhantom isn’t evaluated against any robust aggregation defense.

Minor Weakness:

  1. Results from the pilot study can be shown in the introduction section for easy reading.
  2. The threat model setting can be more detailed. Moreover, key attack settings are missing (trigger design, number of malicious clients, attack rounds, λ, the late-stage schedule, and how “matched budgets” are defined)

[1] Batman: Benign Knowledge Alignment Through Malicious Null Space in Federated Backdoor Attack, CVPR'26.

Questions for authors
  1. Figure 3 (65.35) and Table 2 (88.43) don’t match, though both report donor ASR as 35.42. Which attack does Figure 3 use?
  2. The clean-donor branch is said to be matched by “lineage, defense, and training round”. Which defense is used here? It isn’t described anywhere in the paper.
Reviewer_yPtDOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents an interesting observation that momentum influences subsequent model behavior in algorithms such as FedAvgM. Even when two models have identical weights, different server momentum states can lead to substantially different trajectories of ASR. The authors provide convincing experimental validation of this observation. Building on this finding, they propose DYNAPHANTOM, a backdoor attack in which malicious clients combine historical and current updates to amplify their influence on the optimizer state, thereby increasing backdoor risk.

Critical strengths

The paper offers an interesting perspective and provides convincing evidence that momentum affects post-attack backdoor trajectories by fixing model weights, replacing server momentum, and conducting open-loop replay experiments.

Critical weaknesses

1.The paper does not evaluate DYNAPHANTOM under any update-level defense especially direction similarity based defense. Since DYNAPHANTOM encourages temporally consistent malicious update directions and increases attack strength near the end of the attack, its updates may be distinguishable from benign updates.

2.The experiments in Section 3 examine the effects of replacing the server momentum state in models that have already been backdoored, which is indeed important. But I doubt whether DYNAPHANTOM can actually drive the server model into a backdoored state and steer the server momentum toward the intended state. Given that malicious clients merely combine historical and current updates, how can the authors demonstrate that the submitted updates influence the server momentum in the intended manner, especially when contributions from a large number of benign clients may counteract the malicious contributions?

3.The authors describe DYNAPHANTOM as a persistent backdoor attack. But the experiments only report the ASR over five global rounds on cifar10, and extended to 20 rounds on sst2. Please provide more comprehensive results like how many global rounds it takes for the ASR to first fall below 50% after the attack stops.

4.The experimental evaluation and baseline comparisons are limited. The paper considers only a small number of datasets and models, and compares against author-designed baselines and Neurotoxin, without systematic comparisons with more widely recognized history-aware and persistent backdoor attacks. This feels more like a toy level study.

Questions for authors

see weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.