acceptodds
Under review as a conference paper at ICLR 2027

Detecting Model Poisoning Attacks in Federated Learning via Joint Spatio-Temporal Inspection

Abstract

Detecting model poisoning attacks in federated learning (FL) remains challenging under heterogeneous data distributions and coordinated adversarial behavior. Non-IID data can induce substantial variation among benign updates, leading detectors to misclassify them as malicious, whereas coordinated attackers can submit mutually similar updates to conceal their collective attack behavior. To address these limitations, we propose TSDetector, a joint spatio-temporal model poisoning detection framework that characterizes client behavior from two complementary perspectives. Specifically, from the spatial perspective, TSDetector characterizes the relationships among clients by adaptively fusing an update distance graph which captures discrepancies among client updates with a knowledge discrepancy graph that reflects their similarities in learned knowledge on a reference dataset, from the temporal perspective, TSDetector models client specific update trajectories across communication rounds to identify cross round poisoned attack behaviors compared with normal training dynamics. Extensive experiments across five benchmark datasets and five representative untargeted attacks show that TSDetector reduces average attack impact to 1.16%, with maximum attack induced performance degradation limited to 4.6%. TSDetector avoids severe performance degradation observed in multiple baseline defenses, demonstrating broader applicability across diverse datasets and poisoning strategies.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Detection, Data Heterogeneity

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_z5PJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces TSDETECTOR, a defense against model poisoning in federated learning that jointly inspects spatial and temporal signals. On the spatial side, it analyzes pairwise update distances and discrepancies among client representations within a round; on the temporal side, it models how each client's updates evolve over rounds. Anomaly scores from both views are combined to filter suspicious updates before aggregation. Experiments on five datasets against five attacks report strong detection and defense performance, and component ablations assess the individual contribution of the spatial and temporal modules.

Critical strengths
  • Jointly leveraging within-round client relationships and cross-round client histories is a sensible and appealing design. The two views are sensitive to different attack behaviors, and the ablations provide evidence that they are complementary rather than redundant.

  • The evaluation spans multiple datasets, attack types, and levels of data heterogeneity. Across the settings tested, the results indicate that the combined detector offers effective protection.

Critical weaknesses
  • The paper never reports how TSDETECTOR behaves when no malicious clients are present. The no-attack FedAvg number serves only as a reference and does not show whether turning the defense on degrades normal training. This matters because benign updates under non-IID data can look anomalous. To quantify this cost, the paper should report clean accuracy and the rate at which benign clients are rejected, across different degrees of heterogeneity.

  • The ablation discussion contains numbers that do not match Table 2. The text gives full-model FNRs of 4.33% for PoisonedFL and 1.00% for LIE, while Table 2 lists 3.16% and 3.00%. The authors should state which values are correct and fix the inconsistency.

  • The submitted PDF repeatedly points to appendices that are not included: C.2 for dataset and model details, C.3–C.4 for attack and defense configurations, and C.5 for hyperparameters. Without them, the experimental setup cannot be fully assessed and reproducibility is limited.

Questions for authors
  1. When there are no malicious clients, how does TSDETECTOR compare to FedAvg in terms of accuracy and benign-client rejection rate under varying non-IID settings?

  2. For the full model under PoisonedFL and LIE, which FNR values are correct — those in Table 2 or those in the surrounding text?

  3. Where are the referenced appendices (C.2–C.5)? Please supply the missing experimental details.

2 comments

We greatly appreciate your expert feedback and your insightful concerns. Your concerns and questions are responded as follows:

Question 1: TSDETECTOR behaves when no malicious clients are present?

We thank the reviewer for highlighting the importance of attack-free evaluation. We conducted paired comparisons between TSDetector and FedAvg with 100 benign clients under different degrees of data heterogeneity. Accuracy and FPR are reported in %.

αFedAvg TACCTSDetector TACCTSDetector FPRTACC reduction
0.150.08849.7493.160.339
0.359.54859.3242.290.224
0.561.65660.1411.561.515
0.766.60866.5550.690.053
0.967.05266.6590.900.393

Across these five settings, TSDetector achieves an average FPR of 1.72%, with an average accuracy reduction of only 0.505% relative to FedAvg. Even under strongest heterogeneity setting, α=0.1, the FPR is 3.16%, while accuracy decreases by just 0.339%. These results demonstrate TSDetector's strong performance during attack-free training, preserving accuracy close to FedAvg while maintaining low false-positive rates under heterogeneous data distributions.

We will include these attack-free results in the revision, reporting both accuracy and benign update rejection rates to directly quantify the defense's practical effect on normal training.

Question 2: Which FNR values are correct?

We thank the reviewer for identifying this inconsistency and sincerely apologize for our oversight. Table 2 reports the correct FNRs under the final experimental configuration: 3.16% for PoisonedFL and 3.00% for LIE. After updating the table, we inadvertently retained the earlier values of 4.33% and 1.00% in the text. This reporting error does not affect the final results or the conclusion regarding component complementarity. We will correct the text and verify consistency throughout the revised manuscript.

Question 3: Where are the referenced appendices (C.2-C.5)?

Appendices were submitted with the original submission in the separate supplementary file Appendix.pdf. We sincerely apologize for the unclear location and the resulting inconvenience during review. In the revised manuscript, we will append these materials directly after the main text in a single PDF and verify all cross-references to facilitate access and reproducibility.

AuthorsOpenReview ↗

Anticipating Your Reply as Discussion Deadline Approaching

We would like to express our sincere gratitude for dedicating time to review our paper. We have carefully responded to your questions and made revisions accordingly. As the discussion deadline approaches, we would like to inquire if our responses have adequately addressed your questions and concerns. We are more than willing to address any concerns to ensure a comprehensive resolution and potentially promote your evaluation. Thank you for your time and expertise.

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

This paper proposes TSDetector, a server-side method for identifying malicious updates in federated learning. Its spatial component combines distances between client updates with differences in feature relationships measured on an unlabeled reference set. The temporal component uses CountSketch and a shared LSTM with client-specific states to assign reconstruction-based anomaly scores. An update is rejected if either the spatial or temporal score exceeds its threshold.

Critical strengths
  • The primary strength of the paper is the consistency of the reported results across various attacks. TSDetector achieves an average attack impact of 1.16 percentage points over the 25 settings in Table 1, with a maximum of 4.49 points. In contrast, several baselines perform well against certain attacks but suffer severe degradation against others. TSDetector avoids these large failures in the reported experiments, including under the default non-IID setting with α=0.1\alpha=0.1.

  • The ablation studies provide evidence that the detection components capture different attack behaviors. Removing the temporal branch raises the FNR under PoisonedFL to 94.67%, while removing the knowledge discrepancy graph raises the FNR under LIE to 88.67% (Table 2). These results support the combination of different signals rather than relying on a single anomaly measure. The use of an unlabeled reference set is also a useful feature, particularly compared with defenses that require labeled server data.

Critical weaknesses
  • The claimed contribution of temporal modeling lacks direct empirical validation. In Eqs. (8)–(10), the LSTM processes the current compressed update and reconstructs that same update. While the client-specific recurrent state may contain useful historical information, the paper does not demonstrate whether this information is responsible for improved detection. Removing the entire temporal branch is not an adequate control for this question, as it also removes the reconstruction-based detection capability. This is particularly important given existing methods such as FedREDefense. A memoryless reconstruction model, or an LSTM with client states reset or shuffled across rounds, would provide a much more direct comparison. Without such a test, it remains unclear whether TSDetector learns meaningful temporal behavior or mainly benefits from adding another reconstruction-based anomaly score.

  • The robustness evaluation does not adequately examine the detector's own adaptation process. Algorithm 1 updates client recurrent states and trains the shared LSTM on the 70% of updates with the lowest reconstruction errors before applying the final spatial and temporal filtering rule. Consequently, an update rejected by the spatial detector may have already influenced the temporal model, and the updated client state appears to be retained. This creates a potential vulnerability to gradual contamination that the current attacks do not directly test. The supplementary material also uses a Trimmed-Mean-oriented Fang attack against all defenses, rather than optimizing it specifically for TSDetector. A defense-aware attack targeting both anomaly scores and the online temporal model would be important here. Additionally, the statistical evidence is limited. The experiments report results for seed 42 without cross-seed uncertainty, and there is an inconsistency in the reported FNRs. Table 2 gives FNRs of 3.16% and 3.00% for PoisonedFL and LIE, whereas the discussion reports 4.33% and 1.00%.

Questions for authors
  1. Have you compared the LSTM against a memoryless reconstruction model with similar capacity? What happens if client hidden states are reset or shuffled between clients?

  2. Algorithm 1 updates client states and the shared temporal model before filtering. Is this also the order used in the implementation? Can an update rejected by the spatial detector still contribute to LSTM training, and are rejected clients' updated hidden states retained? How are potentially malicious updates handled during the five-round warm-up?

  3. Have you considered an attacker that explicitly optimizes against both anomaly scores or gradually manipulates the shared temporal model?

  4. Are all reported results based on seed 42, or were additional independent runs conducted? Which FNR values in Table 2 and the accompanying discussion are correct?

  5. In Eq. (2), is the self-neighbor case a=ba=b excluded? The stated index range includes it, but the normalization does not. Is the KL-based knowledge discrepancy symmetrized before constructing the client graph?

2 comments

We greatly appreciate your expert feedback and your insightful concerns. Your concerns and questions are responded as follows:

Question 1 & weakness 1: Comparing the LSTM against a memoryless reconstruction model with similar capacity.

We thank the reviewer for distinguishing the contribution of historical information from that of adding reconstruction-based detection. To directly examine this distinction, we add an LSTM control that resets client states before every round. Before processing current update, each client's hidden and cell states (hi,ci)(h_i,c_i) are reset to initial values rather than carried over from preceding round. It therefore retains reconstruction-based detection while removing cross round carryover of client recurrent states. The table reports FPR/FNR, both in %.

variantLIEFangMin-MaxMin-SumPoisonedFL
Temporal only, client states reset each round4.85/99.332.24/70.004.17/96.675.16/95.334.03/90.00
Temporal only, persistent client states0.27/88.000.45/40.100.35/3.330.29/7.333.88/2.67
TSDetector3.36/3.002.20/0.000.15/0.000.22/0.000.67/3.16

With the same reconstruction-network architecture, resetting states substantially weakens detection. Under Min-Max, Min-Sum, and PoisonedFL, the reset control has FNRs of 96.67%, 95.33%, and 90.00%, respectively. Retaining client historical states reduces these to 3.33%, 7.33%, and 2.67%, while also lowering FPRs. Thus, although Eqs. (8)-(10) reconstruct current update, client history provides useful temporal conditioning for that reconstruction. This control supplies direct evidence for contribution of cross-round states, beyond indirect comparison obtained by removing the entire reconstruction branch.

The standard temporal module nevertheless retains high FNRs under LIE and Fang, showing that temporal evidence alone is not sufficient for every attack. Full TSDetector maintains FNRs no higher than 3.16% across all five attacks, illustrating spatial–temporal complementarity.

Relation to FedREDefense. It's effectiveness involves a specific training-consistency constraint: it optimizes client-specific distilled data-label pairs and reconstructs updates through virtual training, assessing whether submitted updates can be approximately generated by that process. It also retains and refines distilled knowledge across rounds. Its detection mechanism therefore goes beyond simply introducing reconstruction error. TSDetector instead conditions compressed-update reconstruction on client historical states and uses the resulting temporal anomaly evidence to complement spatial detection.

Question 2: (1)Algorithm 1 updates client states and the shared temporal model before filtering. Is this also the order used in the implementation?

As specified in Eqs. (8)-(10) of Section 3.3 and Algorithm 1, current temporal anomaly scores are computed using the previous round's LSTM and decoder parameters. The shared model is subsequently trained on the lowest-error 70%, followed by filtering for aggregation. The implementation follows this order exactly. Thus, an update does not first modify the shared network through training and then get evaluated using the updated network.

(2)Can an update rejected by the spatial detector still contribute to LSTM training, and are rejected clients' updated hidden states retained?

Regarding the specific handling rules, a spatially rejected update still participates in temporal training if it belongs to the low-reconstruction-error training subset. Rejected clients’ historical states are also updated and retained.

(3)How are potentially malicious updates handled during the five-round warm-up?

During the first five rounds, only spatial filtering determines aggregation eligibility, while the temporal module trains and accumulates states without using its scores for rejection. Warm-up does not assume an attack-free initial period.

Question 3: Have you considered an attacker that explicitly optimizes against both anomaly scores or gradually manipulates the shared temporal model?

Our central observation is that appearing benign in one view does not imply passing the checks in other views. Table 2 shows that the temporal only variant has an FNR of 88.00% under LIE, compared with 3.00% for full TSDetector, highlighting the contribution of spatial evidence. We therefore design an adaptive LIE attack using SAS feedback to examine the full detector when the attacker additionally adapts to spatial detection.

The attacker knows current benign client updates and receives its previous round MZ(SAS)\mathrm{MZ}(\mathrm{SAS}), used for filtering. Its malicious update is gadvt=μt+ztσt,g_{\mathrm{adv}}^t=\mu^t+z_t\sigma^t, where μt\mu^t and σt\sigma^t are the mean and standard deviation of benign updates. Starting from zt=1.5z_t=1.5, the attacker reduces its strength the previous spatial score exceeds the threshold and gradually increases it when safely below the threshold.The following table will report TACC, FNR, and FPR across five datasets, all in %.

DatasetTACCFPRFNRAttack impact
MNIST97.4832.0091.700.357
Fashion-MNIST80.7831.0096.651.847
CIFAR-1053.8884.7047.330.862
CIFAR-10027.8070.0049.400.443
Tiny-ImageNet29.2801.3353.830.580

Although the adaptive attack evades detection to some extent, its actual damage remains limited, with an average accuracy reduction of only 0.818 percentage points across the five datasets. This indicates that greater detection evasion does not necessarily translate into stronger attack effectiveness. These results support our central observation: under the evaluated adaptive setting, simultaneously evading spatial and temporal detection while retaining strong attack effectiveness is challenging.

Question 4: (1)Are all reported results based on seed 42, or were additional independent runs conducted?

All results reported in paper use the same random seed, 42. This follows the experimental practice adopted in previous literature. However, to further assess stability and answer this question, we conduct additional experiments with five different seeds (0, 1, 2, 3, and 42) on CIFAR-10 under five attacks. The attack impact are given in the table below, together with the mean and standard deviation, to quantify performance variation across random seeds.

Random seedLIEFangMin-MaxMin-SumPoisonedFL
03.892.013.261.821.31
14.641.971.310.710.43
24.232.512.161.031.68
35.134.090.971.080.29
424.492.841.260.510.26
Mean (standard deviation)4.48 (0.46)2.68 (0.87)1.79 (0.93)1.03 (0.50)0.79 (0.66)

(2)Which FNR values in Table 2 and the accompanying discussion are correct?

We thank the reviewer for identifying this inconsistency and sincerely apologize for our oversight. Table 2 reports the correct FNRs under the final experimental configuration: 3.16% for PoisonedFL and 3.00% for LIE. After updating the table, we inadvertently retained the earlier values of 4.33% and 1.00% in the text. This reporting error does not affect the final results or the conclusion regarding component complementarity. We will correct the text and verify consistency throughout the revised manuscript.

Question 5: In Eq. (2), is the self-neighbor case excluded? The stated index range includes it, but the normalization does not. Is the KL-based knowledge discrepancy symmetrized before constructing the client graph?

First, self-neighbors are excluded in the implementation. For each reference sample bb, we set pb∣b=0p_{b\mid b}=0 and normalize the remaining probabilities only over samples a≠ba\ne b. The index range in Eq. (2) omitted this condition. We will explicitly add it so that the numerator's domain matches the denominator's normalization domain.

Second, the KL-based knowledge discrepancy in Eq. (3) is not symmetrized. The knowledge discrepancy graph is therefore directed and weighted, with Gkt(i,j)≠Gkt(j,i)G_k^t(i,j)\ne G_k^t(j,i) in general. For client ii, this quantity measures the directional discrepancy with its representation-relation distribution as the first argument and client jj's as the second. Following Eq. (6), scoring takes the median of the off-diagonal edge weights in client ii's row of the fused graph, which does not require a symmetric matrix.

Reviewer_A2QwOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes TSDetector, a spatio-temporal detector for untargeted model poisoning in federated learning. It scores each client's update using three signals: the directions of client updates, the relationships among reference samples in client model representations, and the consistency of the client's current update with its history. The first two signals are combined into a spatial anomaly score; a compressed-update reconstruction model provides a temporal anomaly score. Updates exceeding either score's threshold are removed before aggregation. Experiments on five datasets and five attacks compare the method with six defenses, and component ablations examine the signals' complementarity.

Critical strengths

The method combines evidence from update geometry, model behavior on a shared reference set, and client history. The ablation shows that these signals help detect different attacks, making detection more consistent across the evaluated settings.

Critical weaknesses

The defense relies on heuristic scoring and filtering without a theoretical guarantee against strategic attackers. Three potential algorithmic vulnerabilities warrant closer examination:

  • An attacker might keep reconstruction error low enough for its updates to contaminate training of the shared long short-term memory (LSTM) network and multilayer perceptron (MLP) decoder.
  • Changes erased by the fixed CountSketch projection are invisible to the temporal score.
  • Harmful behavior outside the shared reference set's coverage may evade the representation graph.

Theoretical analysis or experiments with adaptive attacks would help clarify how reliably they do so.

Questions for authors
  1. What is the precise technical advance over earlier spatial–temporal poisoning defenses and update-reconstruction detectors, including Fedmvae, FedREDefense, and other relevant prior work? Please identify which parts of the reference-sample graph construction, adaptive graph weighting, and client-specific temporal reconstruction are new, and isolate their contributions against the closest baselines under the same experimental protocol.

  2. How would the detector address the potential vulnerabilities outlined above? In particular, how does it prevent contamination of the temporal model, account for information lost through CountSketch compression, and remain effective when the shared reference set has limited coverage?

  3. Which values for the full method's missed-detection rates in Table 2 are correct? The table reports 3.16% for PoisonedFL and 3.00% for LIE, whereas the accompanying text reports 4.33% and 1.00%.

3 comments

We greatly appreciate your expert feedback and your insightful concerns. Your concerns and questions are responded as follows:

Question 1: What is the precise technical advance over earlier spatial-temporal poisoning defenses and update-reconstruction detectors?

The specific technical advance of TSDetector is to complement parameter geometry with knowledge relations and turn client history into individual temporal anomaly scores, allowing distinct detection criteria to jointly identify malicious updates before aggregation. We clarify differences from Fedmvae and FedREDefense below.

  • Fedmvae: Fedmvae uses temporal information to adjust the aggregate after spatial filtering. TSDetector instead uses client-specific temporal anomaly scores for filtering before aggregation and complements parameter based evidence with representation relations on unlabeled reference inputs.

  • FedREDefense: FedREDefense reconstructs updates through virtual training on client-specific distilled knowledge. TSDetector instead uses recurrent client states to assess temporal consistency of compressed updates, alongside spatial checks of update directions and representation relations.

Relative to these methods, our specific designs are:

  • Knowledge discrepancy graph: It compares relational structures induced by client local models over reference inputs, complementing update direction information, we emphasize this detection construction and its role.

  • Separation based adaptive graph fusion: The weights of update direction and knowledge discrepancy graphs depend on their separation statistics in each round, with a minimum weight preserving both sources. This differs from voting over multiple parameter reconstruction scores.

  • Client level temporal reconstruction and joint filtering: Shared LSTM capture common dynamics, while separate states retain individual histories. Temporal scores directly participate in updates filtering before aggregation.

Question 2: How would the detector address the potential vulnerabilities?

TSDetector combines spatial and temporal evidence because individual detection signals have different blind spots. An update must pass both spatial and temporal checks to enter global aggregation, evading one branch does not directly imply evading entire detector. We address each concern below.

(1)An attacker might keep reconstruction error low enough for its updates to contaminate training of the shared long short-term memory (LSTM) network and multilayer perceptron (MLP) decoder.

We agree that an attacker may maintain low reconstruction error, enter temporal training subset, and influence shared LSTM and decoder. Our method computes current scores using previous round's parameters and then trains temporal model on the lowest-error 70% of updates. This reduces the influence of conspicuous anomalies but does not certify a contamination-free training subset. Spatial rejection also does not automatically exclude an update from temporal model training.

Joint detection remains useful because spatial evidence may prevent malicious updates from entering global aggregation when temporal detection struggles. Table 2 provides concrete evidence. These results support our central empirical finding: behavior that remains inconspicuous in one view may still be identified in another. This demonstrates complementary detection value, without establishing that temporal model cannot be contaminated or guaranteeing robustness against arbitrary adaptive attackers.

(2)Changes erased by the fixed CountSketch projection are invisible to the temporal score.

CountSketch balances the use of update information against the resource costs of temporal modeling. Full model updates are typically high dimensional. Feeding them directly into an LSTM makes its parameters grow with update dimensionality, while full update decoding, gradients, and optimizer states further increase memory and computational costs.

However, CountSketch is used only by temporal branch. The update direction graph operates on updates, while knowledge discrepancy graph examines model knowledge relations induced by those updates. Changes hidden by compression may therefore remain detectable through spatial evidence. This design makes temporal modeling computationally feasible while using complementary criteria to cover potential compression blind spots.

(3)Harmful behavior outside the shared reference set's coverage may evade the representation graph.

TSDetector targets untargeted poisoning that degrades overall model performance. Its knowledge discrepancy graph compares representation relations over shared reference inputs, providing behavioral evidence beyond update geometry without reproducing every client's local data distribution. Experiments with 100 unlabeled samples support its usefulness under evaluated attacks. However, knowledge discrepancy graph may leave some harmful behavior unobserved. TSDetector therefore combines this graph with update direction and temporal evidence to reduce dependence on any single signal.

Question 3: Which values for the full method's missed-detection rates in Table 2 are correct?

We thank the reviewer for identifying this inconsistency and sincerely apologize for our oversight. Table 2 reports the correct FNRs under the final experimental configuration: 3.16% for PoisonedFL and 3.00% for LIE. After updating the table, we inadvertently retained the earlier values of 4.33% and 1.00% in the text. This reporting error does not affect the final results or the conclusion regarding component complementarity. We will correct the text and verify consistency throughout the revised manuscript.

AuthorsOpenReview ↗

Anticipating Your Reply as Discussion Deadline Approaching

We would like to express our sincere gratitude for dedicating time to review our paper. We have carefully responded to your questions and made revisions accordingly. As the discussion deadline approaches, we would like to inquire if our responses have adequately addressed your questions and concerns. We are more than willing to address any concerns to ensure a comprehensive resolution and potentially promote your evaluation. Thank you for your time and expertise.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.