acceptodds
Under review as a conference paper at ICLR 2027

Exploiting Global Model Feedback: An Adaptive Model Poisoning Attack against Robust Federated Learning

Abstract

Federated learning (FL) supports collaborative model training, but its repeated exchange of model states also creates a feedback surface for Model Poisoning Attacks (MPAs). Existing MPAs are largely feedback-agnostic: they predefine malicious updates without assessing how previous updates affected the aggregated model. In this paper, we introduce COMPASS, a feedback-guided MPA that exploits early training dynamics and adaptive aggregation feedback. COMPASS analyzes early broadcast models to extract the dominant training direction and identify Top- important weights. It then crafts an orthogonal perturbation that induces lateral displacement in the optimization trajectory while concentrating attack energy specifically on important weights. Crucially, COMPASS incorporates a matched filter feedback mechanism by embedding a lightweight probe into the malicious update, enabling the attack to dynamically adapt its attack strategy based on the observed aggregation feedback. Evaluations on six datasets, covering 15 aggregation rules and nine baseline attacks, show that COMPASS achieves a mean accuracy reduction of 57.91 percentage points, compared with 23.53 for PoisonedFL, the strongest baseline by this metric (2.46). These findings expose a broader threat to distributed learning systems that repeatedly release aggregated model states.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Model Poisoning Attack, Robust Aggregation

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

This paper proposes COMPASS, an untargeted model poisoning attack against robust federated learning. It estimates a dominant training direction and important parameters from early global model updates, constructs an attack direction orthogonal to the estimated direction, and maintains it between refresh events. A small probe is embedded in malicious updates, and its correlation with subsequent global updates, together with local validation accuracy, guides adaptation of the attack magnitude. The evaluation covers six datasets, nine baseline attacks, and 15 aggregation rules, and reports a larger average accuracy reduction than the evaluated attacks.

Critical strengths
  • The paper studies a relevant security problem: model poisoning without access to benign clients' updates or knowledge of the server's aggregation rule. Exploiting information exposed through global model broadcasts is a meaningful attack setting.

  • The evaluation covers multiple datasets and aggregation mechanisms. The reported improvements over the attack baselines are substantial in several settings, and the appendix includes analyses of probe transmission and feedback control.

Critical weaknesses
  • The positioning relative to PoisonedFL [1] is unclear. Section 3 defines a feedback-guided attack as one that dynamically adjusts its strategy based on global models, but Table 1 marks PoisonedFL as not feedback-guided. PoisonedFL already uses historical global model changes and its own submitted updates to adapt attack magnitudes. The authors should correct Table 1 and explain precisely what information or capability their feedback mechanism adds.

  • The additional insight beyond existing attack designs is not sufficiently established. FedIMP [2] already identifies important parameters using Fisher information and selectively poisons them in an untargeted attack. Neurotoxin [3] exploits the coordinate structure of benign updates to improve attack persistence, while PoisonedFL combines cross-round consistency with dynamic magnitude adjustment. The authors should explain what new insight follows from this combination beyond these existing principles.

  • The breadth of the defense evaluation does not establish sufficiently representative coverage of recent, relevant defenses. Most evaluated methods date from 2017–2022, and the 15 aggregation rules include non-robust Mean. LASA (2025) is an important exception, but COMPASS reduces its accuracy by less than two percentage points on four of the six datasets in Tables 4–9. This limits the conclusion that the attack consistently defeats stronger defenses.

[1] Xie et al. Model Poisoning Attacks to Federated Learning via Multi-Round Consistency. CVPR 2025. [2] Li et al. FedIMP: Parameter Importance-based Model Poisoning Attack against Federated Learning System. Computers & Security, 2024. [3] Zhang et al. Neurotoxin: Durable Backdoors in Federated Learning. ICML 2022.

Questions for authors
  1. Why is PoisonedFL classified as not feedback-guided in Table 1? What capability does COMPASS's feedback provide beyond PoisonedFL's use of historical global model changes?

  2. What is the main additional insight beyond importance-based poisoning, update-structure-aware persistence, and adaptive magnitude control?

  3. Why does COMPASS have limited impact against LASA on four datasets?

1 comment
AuthorsOpenReview ↗

Clarifying COMPASS’s Novelty, Feedback Mechanism, and Experimental Comparisons

The central insight. COMPASS uses broadcast trajectories to construct a persistent lateral attack direction, then uses a matched-filter probe and local validation accuracy (LVA) to adapt attack strength. These observations address where to push the model and whether the push remains visible. All additions and revisions in the manuscript are highlighted in blue.

1. What differs from PoisonedFL? Our PoisonedFL baseline retains its original hypothesis test and adaptive magnitude adjustment (see attackers/poisonedfl.py). PoisonedFL tests whether 50 rounds of global-model changes align with a fixed random direction. It holds the scaling factor on success and multiplies it by 0.70.7 on failure, subject to a lower bound. Thus, its feedback supports holding or shrinking magnitude.

COMPASS instead embeds a small, known signed probe and correlates it with the next global-model change. Its matched filter score (MFS) indicates probe visibility in aggregation, while LVA measures attack effectiveness. Together, they enable increasing, decreasing, or holding magnitude.

2. How do the two signals control magnitude? COMPASS compares feedback trends rather than treating MFS as a binary filtering decision:

  • MFS stable/rising; LVA rising: The probe remains potentially visible while accuracy recovers, suggesting insufficient attack strength. COMPASS gradually increases magnitude.
  • MFS falling; LVA rising: The directional signal weakens while accuracy recovers, suggesting reduced visibility. COMPASS decreases magnitude to avoid stronger filtering.
  • Otherwise: COMPASS holds magnitude when the attack effect is sustained or evidence is inconclusive.

Thus, identical LVA trends can trigger opposite adjustments depending on MFS. Appendix F validates this feedback: counterfactual probe transmission strongly correlates with aggregation acceptance (ρ=0.987\rho=0.987, 64 runs). MFS therefore provides an indication of malicious influence in aggregation.

3. How does COMPASS differ from FedIMP? FedIMP identifies important parameters using Fisher information, which measures loss sensitivity. COMPASS instead uses temporal variance across broadcast global models to identify actively changing parameters that may be effective targets for disrupting convergence. This requires no access to benign clients' updates.

In the new Appendix D (Impact of Attack Target Selection), replacing temporal-variance selection with Fisher-based selection within COMPASS reduces the number of defenses with final accuracy below 20% from 14/14 to 11/14 on MNIST.

Across six datasets and 16 aggregation rules (Tables 2 and 4–11), COMPASS achieves 1.95× the average accuracy impact of FedIMP. This reflects the combined contributions of trajectory-based target selection, orthogonal perturbation, matched-filter feedback, and adaptive magnitude control. The ablation isolates target selection, while the broader comparison evaluates the complete framework.

4. How does COMPASS differ from Neurotoxin? Neurotoxin targets less frequently updated parameters for backdoor persistence, whereas COMPASS selects high-temporal-variance parameters and applies persistent orthogonal perturbations. Neurotoxin is a backdoor attack, while COMPASS is an untargeted model poisoning attack. Given their different objectives, Neurotoxin is outside our experimental scope.

5. Why does COMPASS have limited impact against LASA? LASA retains the Top-kk coordinates with the largest absolute magnitudes, making COMPASS's amplified weights more likely to survive sparsification. However, LASA then applies layer-wise norm and sign filtering. COMPASS amplifies important weights while slightly reducing others to control overall magnitude. If amplified coordinates cause abnormal layer norms, LASA may exclude their malicious contribution. This trade-off helps explain the limited impact on four datasets.

6. How are recent attacks and defenses covered? We thank the reviewer for this suggestion. We added FedIMP (Computers & Security, 2024) as an attack baseline and FoundationFL (NDSS 2025) and ARC (ICLR 2025) as defenses. All three are evaluated across six datasets alongside existing baselines (Tables 2 and 4–9).

Reviewer_qrHvOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes COMPASS, a multi-round model-poisoning attack that uses the broadcast global model as a feedback channel. The stated method estimates a principal direction from early global updates, identifies important coordinates, constructs a Gram–Schmidt-orthogonal malicious update, and embeds a small watermark probe. A matched-filter signal from the next global update is then used by an AIMD controller to adapt the attack magnitude. The paper reports experiments on six image datasets, 15 aggregation/defense rules, and nine baseline attacks, together with ablations for orthogonality, client fraction, heterogeneity, and the matched-filter signal.

Critical strengths
  1. The paper turns the informal intuition that “the global model leaks feedback” into a specific multi-round attack mechanism with an explicit probe and magnitude controller. This makes the security question testable across aggregators rather than tying it to one isolated attack.

  2. The supplementary archive includes an environment specification, runnable configuration files, multiple defenses and baselines, and dedicated MFS/ablation utilities. The reported coverage across datasets and defenses is valuable evidence about where such an attack might matter, even though the code-to-paper mapping must be resolved.

Critical weaknesses
  1. The novelty of COMPASS is not sufficiently established relative to PoisonedFL (Xie et al., CVPR 2025). Specifically, PoisonedFL already leverages historical global model updates to dynamically adjust attack magnitudes and employs hypothesis testing to assess whether recent aggregated updates align with the intended attack direction. Therefore, categorizing PoisonedFL as non-feedback-guided in Table 1 appears questionable and weakens the paper's motivation and novelty claims. Although COMPASS introduces a distinct matched-filter-based feedback mechanism, the authors should more clearly distinguish it from existing feedback-driven strategies and demonstrate its additional benefits.

  2. The ablation study does not adequately isolate the contribution of the proposed matched-filter feedback mechanism. In Appendix F, Static COMPASS disables both adaptive magnitude adjustment and feedback-triggered direction refreshing, making it unclear which component drives the observed improvements. Consequently, the results do not convincingly demonstrate that MFS-based adaptation provides benefits beyond existing feedback strategies. More fine-grained ablations, particularly comparisons against LVA-only adaptation, are necessary to substantiate this central contribution.

Questions for authors
  1. Which exact commit and configuration generated each main table? Please reconcile, line by line, the benign-update access, Captum saliency, centered PCA, test-set evaluation, and watermark scaling with the equations and the broadcast-only threat model. If the archive is an experimental branch, please provide results from a strict implementation of the submitted algorithm.

  2. Does the feedback controller work when the attacker cannot see benign updates, the server test set, the exact malicious fraction, or the defense internals? Please evaluate partial participation and hard-filtering defenses, and quantify the clean-accuracy and detectability cost of the watermark.

2 comments
AuthorsOpenReview ↗

Clarifying COMPASS’s Threat Model, Matched-Filter Feedback, and Robustness

The central clarification. Each of the six main-table datasets has its own configs/FedSGD_<dataset>_config.yaml, ensuring consistent experimental settings across all tasks within the same dataset. The online attacker uses broadcast global models and attacker-owned data, not other information. All additions and revisions in the manuscript are highlighted in blue.

1. How does the code match the stated threat model? The source separates the server's evaluation machinery from the attacker's inputs:

  • Benign updates: Each client first computes its own normal local update. COMPASS then modifies its own update through the non-omniscient entry point; no benign-client update is passed into that method.
  • Captum and PCA: The original source contained an overloaded function that used Captum to select Top-kk weights, but this function was never used in our experiments. We have removed it, retaining only the temporal-variance-based method, which selects Top-kk weights from early broadcast global-model changes. COMPASS then extracts the leading direction from the uncentered update matrix and applies Gram–Schmidt orthogonalization. None of these operations requires benign-client data.
  • Accuracy and hidden server state: LVA evaluates the broadcast model on a held-out subset of attacker-owned data. The server's test set is used for reported accuracy, not supplied to COMPASS. Magnitude adaptation reads MFS and LVA trends; it does not query the exact malicious fraction or the aggregation rule's internal selection, clipping, or weights.

2. What does MFS add beyond prior feedback and LVA? We acknowledge that PoisonedFL already uses global-model feedback and have corrected Table 1. PoisonedFL tests whether 50 rounds of global-model changes align with a fixed random direction, holding its scaling factor when the test succeeds and reducing it by a factor of 0.70.7 otherwise. COMPASS instead embeds a small, known signed probe and correlates it with the next global-model change. MFS provides a noisy indication of probe visibility in aggregation, while LVA measures attack effectiveness. Their combination enables increasing, decreasing, or holding attack magnitude.

3. How do the two signals guide adaptation? COMPASS compares recent MFS and LVA trends:

  • MFS stable/rising; LVA rising: The probe remains potentially visible while accuracy recovers, suggesting insufficient attack strength. COMPASS gradually increases magnitude.

  • MFS falling; LVA rising: The directional signal weakens while accuracy recovers, suggesting reduced probe visibility. COMPASS decreases magnitude to avoid stronger filtering.

  • Otherwise: COMPASS holds magnitude when the attack effect is sustained or the feedback is inconclusive.

Thus, identical LVA trends can trigger opposite adjustments depending on MFS. Appendix F validates this feedback: counterfactual probe transmission strongly correlates with aggregation acceptance (ρ=0.987\rho=0.987, 64 runs). MFS can therefore provide an indication of the extent to which malicious updates are incorporated into the aggregation result.

4. Does joint feedback improve attack effectiveness? We provide two complementary results:

  • Behavior — new Fig. 7 (Appendix E): Starting from λ0=3\lambda_0=3, magnitudes diverge across 16 defenses. RFA rises substantially, whereas Krum remains near its initial magnitude, illustrating defense-dependent adaptation.
  • Outcome — new Table 4: We compare Full COMPASS, MFS-only, LVA-only, fixed magnitudes, and bounded ranges while keeping all other mechanisms, including orthogonal direction selection and refresh, identical. Across 14 MNIST defenses, Full achieves 9.8–11.4% accuracy under all 14, whereas MFS-only and LVA-only each succeed against only five. Fixed magnitudes perform inconsistently across Krum and CenteredClipping, while Full succeeds against both, demonstrating the benefit of joint MFS–LVA feedback.

5. Does the attack survive restricted participation and strong filtering? In the partial-participation evaluation (Appendix B), only 10 of 100 clients participate per round. Across 15 defenses on MNIST, COMPASS ends at 9.80–10.09% test accuracy, compared with 10.03–11.40% under full participation. The evaluated reduction in participation therefore does not erase the attack.

6. What is the watermark’s detectability cost? The injected probe is zt=ηtPtz_t=\eta_tP_t on Top-kk coordinates, with ∣zt∣2/∣utmal∣2=10−4|z_t|_2/|u_t^{\rm mal}|_2=10^{-4}. Its magnitude is 10,000 times smaller than the malicious update, making its impact on training negligible. The next broadcast update is correlated with that same injected probe to obtain MFS. We have added a new appendix section, “I. Impact of the Matched-Filter Probe,” to evaluate its effect on benign training. As shown in Table 18, the mean absolute accuracy difference across 12 dataset–aggregator pairs is only 0.62 percentage points, with no consistent or systematic degradation in test accuracy.

Thank you for the detailed response and clarifications. After carefully considering your rebuttal together with the comments from the other reviewers, I have decided to keep my current score. I appreciate the authors’ efforts in addressing the concerns raised.

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

The paper presents COMPASS, an untargeted model-poisoning attack against robust federated learning. It estimates a dominant training direction and a set of important parameters from early global models, then constructs perturbations orthogonal to that direction. A small embedded probe provides feedback about aggregation, which is used alongside local validation accuracy to adapt the attack. The evaluation covers six datasets, 15 aggregation rules, and nine baseline attacks.

Critical strengths

The strongest aspect is the breadth and size of the reported improvements. COMPASS produces an average accuracy reduction of 57.91 percentage points, compared with 23.53 for PoisonedFL, and has the highest dataset-averaged impact under 13 of the 15 aggregation rules. Although effectiveness varies across datasets, these results suggest a substantial vulnerability worth investigating.

I also appreciate the effort to examine the mechanism. The appendix goes beyond correlations between probe scores and accuracy by replaying aggregation with controlled perturbations. The paired trajectory analysis and orthogonal-versus-reverse comparison provide additional evidence about how the attack changes training. These analyses make the paper more informative than an attack evaluated only through final accuracy.

Critical weaknesses

My main concern is the treatment of prior work. Table 1 labels PoisonedFL as not feedback-guided, and the paper repeatedly contrasts COMPASS with supposedly feedback-agnostic attacks. However, PoisonedFL already adjusts its magnitude using previous global-model changes and hypothesis testing to assess whether its perturbations survive aggregation. This directly weakens the stated novelty argument. The potentially new contribution is the particular probe-based feedback and structured attack direction, rather than using global-model feedback itself. It also raises an important question about whether the evaluated PoisonedFL implementation retains its original adaptive mechanism.

The experiments do not yet cleanly isolate the added value of the matched-filter feedback. The static-versus-full comparison disables both magnitude adaptation and direction refresh, so the improvement cannot be attributed specifically to the probe. There is no corresponding adaptive controller using local validation accuracy alone. Moreover, this ablation covers only MNIST and three seeds; feedback helps against some defenses, has little effect against others, and worsens results against LASA. The extensive probe validation establishes that a signal can be measured, but its additional benefit for controlling the attack remains less firmly established. A related limitation is that attack parameters are tuned separately for every dataset–aggregation pair using up to 100 trials. Equal search budgets are useful for comparison, but this is a stronger evaluation setting than encountering an unknown defense in a single training run. The distinction matters when interpreting the claim that COMPASS is aggregation-agnostic

Questions for authors

Does the PoisonedFL baseline retain its original global-feedback-based magnitude adaptation and hypothesis test? How do the authors reconcile that mechanism with its classification as not feedback-guided in Table 1?

Are the 100 tuning trials performed on separate training runs, and are the selected parameters evaluated on independent seeds? What performance is obtained with one configuration transferred across aggregation rules?

Can the authors compare the full controller with an LVA-only adaptive controller while keeping direction refresh identical? This would help establish whether the matched-filter signal adds value beyond the other feedback already available.

1 comment
AuthorsOpenReview ↗

Clarifying COMPASS's Novelty and Validating Feedback Adaptation

The central distinction. We acknowledge that PoisonedFL already uses global-model feedback and have corrected Table 1. COMPASS's novelty lies in its matched-filter probe and joint MFS–LVA feedback, rather than feedback itself. All additions and revisions in the manuscript are highlighted in blue.

1. What differs from PoisonedFL? We confirm that our PoisonedFL baseline retains its original hypothesis test and adaptive magnitude adjustment (see attackers/poisonedfl.py in the supplementary code). PoisonedFL tests whether 50 rounds of global-model changes align with a fixed random direction. Its implementation holds the scaling factor when the test succeeds and multiplies it by 0.70.7 on failure, subject to a lower bound. Thus, its feedback supports holding or shrinking magnitude.

COMPASS instead embeds a small, known signed probe and correlates it with the next global-model change. Its matched filter score (MFS) provides a noisy indication of probe visibility in aggregation, while local validation accuracy (LVA) measures attack effectiveness. Their combination enables increasing, decreasing, or holding magnitude.

2. How do the two signals control magnitude? COMPASS compares recent feedback trends rather than treating MFS as a binary filtering decision:

  • MFS stable/rising; LVA rising: The probe remains potentially visible while accuracy recovers, suggesting insufficient attack strength. COMPASS gradually increases magnitude.
  • MFS falling; LVA rising: The directional signal weakens while accuracy recovers, suggesting reduced probe visibility. COMPASS decreases magnitude to avoid stronger filtering.
  • Otherwise: The attack effect is sustained or evidence is inconclusive. COMPASS holds magnitude.

Thus, identical LVA trends can trigger opposite adjustments depending on MFS. Appendix F validates this feedback: counterfactual probe transmission strongly correlates with aggregation acceptance (ρ=0.987\rho=0.987, 64 runs). MFS can therefore provide an indication of the extent to which malicious updates are incorporated into the aggregation result.

3. Does matched-filter feedback improve effectiveness? We provide two complementary results:

  • Behavior — new Fig. 7 (Appendix E): Starting from λ0=3\lambda_0=3, magnitudes diverge across 16 defenses. RFA rises substantially, whereas Krum remains near its initial magnitude, illustrating defense-dependent adaptation.
  • Outcome — new Table 4: We compare Full COMPASS, MFS-only, LVA-only, fixed magnitudes, and bounded ranges while keeping all other mechanisms, including orthogonal direction selection and refresh, identical. Across 14 MNIST defenses, Full achieves 9.8–11.4% accuracy under all 14, whereas MFS-only and LVA-only each succeed against only five. Fixed magnitudes perform inconsistently across Krum and CenteredClipping, while Full succeeds against both, demonstrating the benefit of joint MFS–LVA feedback.
  • Why can feedback perform worse against LASA? LASA retains each client's largest-magnitude Top-kk weights before filtering updates based on layer-wise norms and directions. Amplifying selected weights makes them more likely to survive Top-kk selection, but excessive amplification can make their layer norms deviate from benign updates, triggering layer-wise filtering. This trade-off explains why increasing attack magnitude can sometimes reduce effectiveness against LASA.

4. How were hyperparameters selected? The 100 trials represent the maximum preliminary tuning budget, not 100 trials necessarily performed per defense. Each trial is an independent training run using seed 42 and lasts only 200 rounds, rather than the 1,000 rounds used in formal evaluation.

Like PoisonedFL, whose implementation provides multiple candidate initial magnitudes, COMPASS requires a reasonable initialization. Preliminary experiments identified a unified configuration: λ0=3\lambda_0=3, mI=100m_I=100, mU=0.99m_U=0.99. This configuration is applied unchanged across all 16 defenses. Final evaluation uses five seeds (42–46), including four (43–46) distinct from the tuning seed.

New Fig. 7 demonstrates that this single configuration, with λ0=3\lambda_0=3, compromises all 16 defenses without defense-specific adjustments during evaluation. Importantly, offline tuning determines the common initialization; during training, COMPASS adapts magnitude using attacker-observable MFS and LVA without knowing the aggregation rule's internal mechanisms. Our aggregation-agnostic claim concerns this defense-independent online adaptation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.