acceptodds
Under review as a conference paper at ICLR 2027

Morpheus: Modality-Agnostic Adversarial Attack from Leaked Models in Federated Learning

Abstract

Federated learning (FL) creates a distinctive attack surface by repeatedly exposing intermediate global models to participating clients. A malicious client can retain one such checkpoint as a surrogate and use it to craft adversarial examples (AEs) against the final deployed model, even though the model may undergo subsequent training. In this paper, we investigate an intriguing property, which we term temporal transferability: the ability of an AE crafted against an early checkpoint along a training trajectory to remain effective against the converged model. Although extensive prior work has studied transfer attacks, where AEs crafted against one model are transferred to another, existing approaches are often tailored to specific modalities and do not explicitly account for model evolution over time. We identify two fundamental ingredients that govern temporal transferability: (i) an AE should achieve a sufficiently large adversarial margin, and (ii) it should lie in a region of the leaked model's feature space that is insensitive to parameter changes, making it more likely to remain on the adversarial side of the decision boundary as training progresses. Motivated by these observations, we introduce the weight-space margin (WSM), which quantifies an AE's robustness to changes in model parameters and can be computed using only the leaked checkpoint. Building on this insight, we propose Morpheus, the first modality-agnostic temporal transfer attack. Morpheus optimizes each AE across randomly sampled subnetworks of the leaked checkpoint, reducing its dependence on the exact leaked weights while requiring only the checkpoint and the inputs to be attacked. Across three modalities, six datasets, and six architectures, Morpheus consistently outperforms the strongest of six baseline attacks. When the checkpoint is leaked after only 30% of training, Morpheus improves the attack success rate over the strongest baseline by up to 20.00%.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
federated learning, adversarial attacks, adversarial examples, adversarial transferability, model leakage, temporal transfer, weight-space smoothing

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_8HNdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates temporal transferability in Federated Learning, where adversarial examples crafted on an early leaked checkpoint remain effective against the final deployed model. The authors propose Morpheus, a modality-agnostic attack that optimizes perturbations across stochastic subnetworks of the leaked source model to achieve weight-space smoothing. Experiments across image, audio, and point cloud modalities demonstrate that Morpheus outperforms existing baselines.

Critical strengths
  • The paper successfully extends temporal transfer attacks beyond vision by operating purely in weight space via stochastic subnetwork sampling, validating its effectiveness across image, audio, and point cloud tasks. This addresses a significant gap in FL security literature where prior work was limited to image-specific transformations.

  • The evaluation is comprehensive, covering 12 dataset-architecture pairs and demonstrating consistent superiority over six baselines. The introduction of the Weight-Space Margin provides a useful diagnostic tool that correlates strongly with empirical transfer success.

Critical weaknesses
  • Both the concept and the mechanism of Morpheus appear closely related to existing work [R1-R3]. The term "temporal transferability" seems conceptually equivalent to the "Cross-Stage Adversarial Transferability" of Kim & Lee (2025), which is not cited. Proteus, in turn, appears to share the same threat model and to be the strongest relevant baseline on the image pairs, yet it is absent from the main comparison table. This makes it difficult to assess whether Morpheus offers a genuine algorithmic improvement over the state of the art, or mainly generalizes existing mechanisms to modalities where they had not yet been applied.

  • WSM rests on a Markov inequality bound derived for a fixed parameter perturbation distribution Q_ρ. As Appendix B.1 acknowledges, this bound makes no statement about the sequential training drift that the paper sets out to explain, so it does not appear to cover the phenomenon it is meant to motivate. Related notions of margin and sharpness under weight perturbations have already been characterized [R4-R5], including a transferability metric based on output change under surrogate weight perturbation in AWT, which is also a baseline here. It would help to clarify why WSM is necessary and how it adds predictive power beyond these static measures in capturing temporal evolution.

[R1] Exploring Cross-Stage Adversarial Transferability in Class-Incremental Continual Learning. IEEE MMSP 2025, arXiv:2508.08920.

[R2] Proteus: Model Leakage-Induced Adversarial Attack in Federated Learning. ECCV 2026.

[R3] Enhancing Adversarial Transferability with Checkpoints of a Single Model's Training. CVPR 2025.

[R4] Formalizing Generalization and Robustness of Neural Networks to Weight Perturbations. NeurIPS 2021, arXiv:2103.02200.

[R5] Enhancing Adversarial Transferability with Adversarial Weight Tuning. AAAI 2025, arXiv:2408.09469.

Questions for authors

Q1: Can you include Proteus (ECCV 2026) in the main comparison table for all image pairs? As the most closely related work with the same threat model, its performance is critical for establishing the true incremental gain of Morpheus.

Q2: On the diagnostic set, how does WSM compare with the adversarial margin and with sharpness taken separately for ranking examples by transfer outcome? This would indicate whether WSM carries predictive information beyond its two ingredients.

Q3 : Is there any bound or source-side analysis that relates WSM to survival under actual training drift, rather than under draws from Q_ρ?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 8HNd Part 1

We thank Reviewer 8HNd for the insights and references.

Q1 / W1: Can Proteus be included in the main comparison for all image pairs?

Morpheus remains highly competitive with Proteus even when Proteus's original transformations are enabled. The results below confirm that, even without transformations, Morpheus remains competitive with, and can even surpass, Proteus with transformations. We will include Proteus in our main results.

ASR (%) comparison with Proteus, the state-of-the-art modality-specific attack (Morpheus with N=32):

AttackRAF-DB/ResNet18RAF-DB/VGG16TinyImageNet/ResNet18TinyImageNet/VGG16
Proteus with transformations77.9654.2788.1587.01
Morpheus without transformations82.8056.2887.2085.71

W1 ([R1], [R3]): How does temporal transferability in FL relate to [R1] and [R3]?

Temporal transferability in FL differs from the cross-stage transferability studied in [R1]. [R1] studies class-incremental continual learning, where a centrally trained model is updated on a known sequence of tasks. In FL, the drift between the leaked round and deployment results from aggregating updates from randomly sampled clients whose data the attacker never sees. The attacker has only one checkpoint, with no training data or feedback from the deployed model.

As shown in the table below, under the FL threat model, Morpheus outperforms the strongest baseline from [R1] and [R3] by approximately 14 to 36 percentage points across the four dataset/model settings. We evaluated [R1] and [R3] under this threat model at ε = 8/255, using the official attack settings for [R1]. With only one leaked checkpoint, the checkpoint ensemble of [R3] reduces to I-FGSM-10 on the leaked model.

ASR (%) comparison across attacks:

AttackRAF-DB/ResNet18RAF-DB/VGG16TinyImageNet/ResNet18TinyImageNet/VGG16
[R1] FGSM50.0020.1072.9956.71
[R1] PGD-1030.1111.0660.1936.80
[R1] AutoAttack23.9214.0732.7012.55
[R3] Single leaked checkpoint37.9015.8369.6735.50
Morpheus (Ours)82.8056.2887.2085.71

We will refine the related work section to include [R1] and [R3].

AuthorsOpenReview ↗

Response to Reviewer 8HNd Part 2

Q2 / W2: Does WSM carry predictive information beyond margin and sharpness alone?

WSM provides stronger predictive information overall than either of its ingredients alone. This holds for ranking individual AEs as well as for ranking attacks. On the diagnostic set of Fig. 3, ranking the 1,593 AEs of its five attacks by transfer outcome gives a ROC AUC of 0.68 for WSM, against 0.58 for margin only and 0.55 for negative sharpness only. Across the 12 dataset/model settings, the same per-example AUC averages 0.81 for WSM, against 0.74 and 0.59, and WSM is the highest in every setting.

For ranking attacks, we use the Spearman correlation between the mean of each score over an attack's AEs (WSM, margin only, or negative sharpness only) and its transfer ASR. A stronger positive correlation indicates better agreement with the ranking by transfer ASR. The table below reports the results across all 12 dataset/model settings.

Dataset/Model SettingWSM (Ours)Margin only−Sharpness only
RAF-DB / ResNet180.710.68−0.39
RAF-DB / VGG160.710.210.21
TinyImageNet / ResNet180.960.640.00
TinyImageNet / VGG160.820.29−0.04
CREMA-D / BC-ResNet80.890.75−0.64
CREMA-D / CBAM-DenseNet1210.860.71−0.43
UrbanSound8K / BC-ResNet80.990.94−0.59
UrbanSound8K / CBAM-DenseNet1210.960.890.29
ModelNet40 / PointNet1.000.89−0.61
ModelNet40 / PointMLP0.960.86−0.86
ShapeNetPart / PointNet1.000.960.89
ShapeNetPart / PointMLP0.890.960.96
Mean over 12 settings0.900.73−0.10

Overall, WSM achieves a mean Spearman correlation of 0.90, substantially higher than either margin or negative sharpness alone. This provides strong evidence supporting its design.

WSM also differs from [R4] and [R5] in what it measures. [R4] derives worst-case bounds on the class margin of a model under weight perturbations and uses them to train robust models. It does not score individual AEs. AWT [R5] proposes a transferability score that measures only the output change of the surrogate on AEs under weight perturbations. WSM is a per-AE score computed on the leaked checkpoint, defined as the log-ratio of the adversarial margin to its change under weight perturbations. The results above show that this change alone, without the margin, is the weakest of the three scores on average.

Q3 / W2: Is WSM linked to attack survival under actual training drift?

WSM is empirically associated with attack survival under actual training drift across the 12 real FL trajectories. In FL, the drift between leakage and deployment results from aggregated updates of clients the attacker never sees, so it cannot be fully predicted from the leaked checkpoint. WSM is therefore a static, source-only diagnostic of how reliably an AE survives unknown drift, not a predictor of the drift.

The Markov bound in App. B.1 is stated for the perturbation law Q_ρ and motivates the log-ratio form of WSM. A bound for the actual drift would need assumptions on the updates of the unseen clients. The relation of WSM to the temporal evolution of the model is therefore tested on the outcome of real training. WSM computed on the leaked checkpoint predicts the ranking of attacks by transfer ASR at deployment with a mean Spearman correlation of 0.90, and ranks individual AEs with a mean ROC AUC of 0.81 (see Q2).

Thank you once again for your careful comments. As the discussion phase is approaching its end, we would like to kindly confirm whether our response has sufficiently addressed your concerns, including the comparison with Proteus, the discussion of the related work you pointed out, and the questions on WSM. Should any points still require clarification, we would be glad to provide further details. If you are satisfied with our responses, we would greatly appreciate your consideration in adjusting the evaluation scores accordingly. We sincerely look forward to your feedback.

Reviewer_z8puOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates temporal transferability in federated learning (FL): whether adversarial examples generated from an early leaked global checkpoint remain effective against the final deployed model. It introduces the weight-space margin (WSM), which combines adversarial margin and sensitivity to parameter changes to characterize transferability. Building on this motivation, the paper proposes Morpheus, modality-agnostic adversarial attack by optimizing over stochastic subnetworks of the leaked checkpoint using multi-rate connection dropping and aggregated gradients Experiments across image, audio, and point-cloud modalities evaluate temporal transferability under different settings show that Morpheus outperforms other attacks.

Critical strengths

The paper addresses an important problem with a clear threat model: how leaked checkpoints can be exploited to generate adversarial examples to fool the final model in FL.

The design is well-motivated and intuitive: a large adversarial margin and insensitivity to parameter changes over training.

The experiments are detailed and comprehensive. The evaluation for the temporal transferability of adversarial attacks across multiple modalities under different settings show that Morpheus outperforms other attacks.

Critical weaknesses

The disabling of modality-specific attacks requiring specific input transformations (like Proteus or RaPA in images) makes the comparison of image baselines potentially unfair.

Morpheus uses 32 subnetworks per iteration, while the others like PGD seem use only 1, and RaPA uses 5, which introduces extra computational cost. The paper does not discuss or analyze this cost.

Providing more failure cases and finer-grained cases (which modality/defense fail this attack more often) could still strengthen this paper.

Questions for authors

Does the attack still remain effective under point-could or audio modalities with defenses (Gaussian Input Randomization, Randomized Smoothing)?

Does Morpheus's modality-agnostic design compromise its effectiveness compared to modality-specific attacks such as Proteus and RaPA with their original image transformations enabled?

Could the authors discuss the extra cost of subnetwork sampling in Morpheus?

2 comments
AuthorsOpenReview ↗

Response to Reviewer z8pu

We thank Reviewer z8pu for the positive and helpful comments.

Q1 / W3: Does Morpheus remain effective on point clouds and audio under defenses?

Morpheus is resilient to defenses on point clouds and audio. As shown in the tables below, its ASR decreases only from 87.92% to 84.76% under Gaussian input randomization and to 79.54% under randomized smoothing (RS) on point clouds. These defenses are even less effective on audio, reducing Morpheus's ASR from 91.12% to 88.32%. Morpheus maintains the highest ASR under both defenses across both modalities.

ASR (%) on point clouds, ModelNet40/PointNet:

MethodNo defenseGaussianRS
FGSM26.1232.9426.80
PGD61.8043.7038.90
MI-FGSM77.5370.8465.71
NI-FGSM36.2433.7930.26
AWT12.0821.6917.29
RaPA75.9369.4563.34
Morpheus (Ours)87.9284.7679.54

ASR (%) on audio, UrbanSound8K/CBAM-DenseNet121:

MethodNo defenseGaussianRS
FGSM40.5839.3339.42
PGD65.0859.5459.56
MI-FGSM67.6963.2863.28
NI-FGSM60.3456.9757.02
AWT24.0415.1115.23
RaPA74.2772.4572.59
Morpheus (Ours)91.1288.3288.32

Q2 / W1: How does Morpheus compare with Proteus and RaPA when their image transformations are enabled?

Morpheus remains highly competitive with, and often surpasses, modality-specific attacks with their original transformations enabled. In the comparison with RaPA below, we match N across attack variants. These results demonstrate that Morpheus's modality-agnostic design delivers competitive performance even without relying on modality-specific input transformations.

ASR (%) comparison with RaPA (N=5):

AttackRAF-DB/ResNet18RAF-DB/VGG16TinyImageNet/ResNet18TinyImageNet/VGG16
RaPA with transformations61.2960.5578.6772.73
Morpheus without transformations80.9152.7687.2084.42

ASR (%) comparison with Proteus, the state-of-the-art modality-specific attack (Morpheus with N=32):

AttackRAF-DB/ResNet18RAF-DB/VGG16TinyImageNet/ResNet18TinyImageNet/VGG16
Proteus with transformations77.9654.2788.1587.01
Morpheus without transformations82.8056.2887.2085.71

Q3 / W2: What is the extra computational cost of subnetwork sampling?

Morpheus's effectiveness gains are not simply due to a higher computational cost. Subnetwork sampling requires 640 gradient evaluations per AE at the default setting of N=32. This is fewer than the approximately 800 required by Proteus, against which Morpheus achieves competitive or superior performance, as shown in Q2. For RaPA, the table under Q2 shows that, even after matching N, Morpheus without transformations achieves large gains in three of the four settings. For PGD and MI-FGSM, which each use 20 gradient evaluations per AE by default, our new experiments show that increasing their budgets to 320 or 640 (matching Morpheus) does not close the gap:

ASR (%) with different gradient evaluation budgets:

Attack, gradient evaluations per AERAF-DB/ResNet18RAF-DB/VGG16TinyImageNet/ResNet18TinyImageNet/VGG16
PGD, 20 / 320 / 64050.54 / 51.08 / 52.9625.63 / 25.63 / 24.8778.20 / 77.25 / 77.2553.25 / 58.87 / 61.47
MI-FGSM, 20 / 320 / 64055.91 / 55.38 / 55.3828.64 / 30.90 / 30.1581.04 / 80.09 / 79.6263.20 / 63.64 / 65.37
Morpheus, 640 (Table 4)82.8056.2887.2085.71

Thank you for the detailed response and additional experiments. The evaluations on point clouds and audio, the transformations with Proteus and RaPA, and the computational cost comparisons have effectively addressed my concerns. However, the relevance of WSM to training trajectories remains unclear, as pointed out by another reviewer. Therefore, I maintain my original rating.

Reviewer_xVWWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether adversarial examples generated from an early FL checkpoint can fool the final deployed model. Morpheus optimizes perturbations over randomly pruned subnetworks, sampling pruning rates from a log-uniform distribution. The paper also introduces the weight-space margin (WSM) to assess temporal transferability. Experiments cover three modalities and twelve dataset–architecture pairs.

Critical strengths

1. The threat model is practical and clearly defined. Participating clients receive intermediate checkpoints, and the attack requires neither additional training data nor target-model feedback.

2. The method is simple and applies across modalities without designing separate input transformations.

3. The evaluation covers multiple architectures and leakage stages, with substantial improvements in several settings.

Critical weaknesses

1. My main concern is what accounts for the improvement over existing attacks. RaPA is evaluated without its image transformations even on image tasks, while its pruning scope and sampling budget also differ from Morpheus. A common cross-modal protocol is reasonable, but leaves performance against the complete image baseline unanswered. Proteus appears in Figure 3 and Table 2, but not in the main comparison. I appreciate the lower-budget result in Appendix D.6 and the rate-sampling ablation in Table 1. What remains unclear is how much of the gain over RaPA comes from the broader pruning scope rather than multi-rate sampling.

2. The main WSM evidence needs closer alignment with its definition. Equation (3) defines a log-ratio, whereas Equation (20) standardizes the two log terms separately. This changes their relative weighting and can change rankings, so Figure 3 does not directly validate the original score. Figure 5 and Appendix B.2 offer supporting evidence, but do not establish its predictive advantage over margin or sharpness alone across modalities. The bound itself concerns perturbations drawn from QρQ_\rho, not subsequent training drift, as the authors acknowledge.

3. I could not reconcile the point-cloud numbers with the stated metric without additional information. If Table 4 reports single-run values from Equation (28), each ASR should have the form 100k/n100k/n, with integer success count kk and a shared target-clean denominator n≤512n\leq512. Under these conditions, the ModelNet40/PointNet entries 26.1026.10% and 77.5077.50% cannot both result from direct rounding to two decimal places. Please provide the counts and explain any averaging or intermediate rounding.

4. The main results do not report variation across independently trained FL trajectories. Attack-seed repetitions on fixed checkpoints do not capture changes in initialization, data partitioning, or client sampling. A few independent training runs on representative settings would help establish whether the reported gains persist.

Overall Assessment

The problem is worthwhile, and the method shows promise. However, the comparisons leave the incremental contribution unclear, and the evidence for WSM is not yet convincing enough. The point-cloud reporting issue also needs clarification. I would consider raising my score if the response resolves these concerns with clear explanations and supporting evidence.

Questions for authors

1. Under matched gradient budgets, how much of the improvement over RaPA comes from broader pruning scope, and how much from multi-rate sampling? How does the complete RaPA configuration perform on image tasks?

2. Does raw WSM reproduce Figure 3’s trends on representative image, audio, and point-cloud settings, and does it outperform either component alone as a predictor?

3. What counts and averaging procedures underlie the point-cloud results? Can the authors also report variability across independent FL training runs for a few representative settings?

4 comments
AuthorsOpenReview ↗

Response to Reviewer xVWW Part 1

We thank Reviewer xVWW for the time and effort.

Q1 / W1: Where does the gain over RaPA come from, and how does the complete RaPA perform on images?

At a matched budget of 100 gradient evaluations per AE, Morpheus outperforms RaPA across all four dataset/model settings without RaPA's original transformations (RDI+TI), and in three of the four settings with these transformations enabled. With log-uniform sampling fixed, changing the drop set from all Conv2d and Linear weights to batch-normalization and linear parameters alone lowers ASR from 80.91% to 73.92% on RAF-DB/ResNet18 (6.99 percentage points) and from 87.20% to 84.83% on TinyImageNet/ResNet18 (2.37 percentage points). Replacing log-uniform sampling with a fixed rate equal to its mean (0.0195) yields 78.76% on RAF-DB/ResNet18 (2.15 percentage points lower) and leaves TinyImageNet/ResNet18 unchanged at 87.20%.

The table below compares the ASR (%) of two RaPA variants (without transformations and the complete configuration with transformations) with Morpheus. We use a matched budget of 100 gradient evaluations per AE for all attacks (N=5). The complete configuration improves RaPA only on RAF-DB/VGG16 and lowers its ASR by 3.95 to 8.87 percentage points on the other three dataset/model settings. Even without transformations, Morpheus remains highly competitive with, and often surpasses, both RaPA variants.

AttackRAF-DB/ResNet18RAF-DB/VGG16TinyImageNet/ResNet18TinyImageNet/VGG16
RaPA (no transformations)70.1646.2382.6276.91
RaPA (complete configuration)61.2960.5578.6772.73
Morpheus (Ours)80.9152.7687.2084.42

W1 (Proteus): How does Morpheus compare with Proteus in the main comparison?

Morpheus remains highly competitive with Proteus even when Proteus's original transformations are enabled. The results below confirm that, even without transformations, Morpheus remains competitive with, and can even surpass, Proteus with transformations. We will include Proteus in our main results.

ASR (%) comparison with Proteus, the state-of-the-art modality-specific attack (Morpheus with N=32):

AttackRAF-DB/ResNet18RAF-DB/VGG16TinyImageNet/ResNet18TinyImageNet/VGG16
Proteus with transformations77.9654.2788.1587.01
Morpheus without transformations82.8056.2887.2085.71

Q2 / W2: Does raw WSM reproduce the trends of Fig. 3 and outperform margin or sharpness alone?

WSM predicts transfer ASR better overall than either component alone. Specifically, we assess this using the Spearman correlation between each score (WSM, margin only, or negative sharpness only) and transfer ASR. A stronger positive correlation indicates better agreement with the ranking by transfer ASR. The table below reports the results across all 12 dataset/model settings.

Dataset/Model SettingWSM (Ours)Margin only−Sharpness only
RAF-DB / ResNet180.710.68−0.39
RAF-DB / VGG160.710.210.21
TinyImageNet / ResNet180.960.640.00
TinyImageNet / VGG160.820.29−0.04
CREMA-D / BC-ResNet80.890.75−0.64
CREMA-D / CBAM-DenseNet1210.860.71−0.43
UrbanSound8K / BC-ResNet80.990.94−0.59
UrbanSound8K / CBAM-DenseNet1210.960.890.29
ModelNet40 / PointNet1.000.89−0.61
ModelNet40 / PointMLP0.960.86−0.86
ShapeNetPart / PointNet1.000.960.89
ShapeNetPart / PointMLP0.890.960.96
Mean over 12 settings0.900.73−0.10

Overall, WSM achieves a mean Spearman correlation of 0.90, substantially higher than either margin or negative sharpness alone. This provides strong evidence supporting its design.

Furthermore, with raw WSM, we observe the same trends as in Fig. 3 across all three modalities. On images, for the five attack families in Fig. 3 (PGD, AWT, RaPA, Proteus, and Morpheus), WSM achieves Spearman correlations of 1.00 on RAF-DB/ResNet18, 1.00 on RAF-DB/VGG16, 0.90 on TinyImageNet/ResNet18, and 0.97 on TinyImageNet/VGG16, compared with margin-only correlations of 0.30, −0.10, 0.30, and −0.21, respectively. On audio and point clouds, in the analysis corresponding to panel (c), transfer ASR rises from 70.8% in the lowest-WSM group to 100% in the highest-WSM group on UrbanSound8K/CBAM-DenseNet121 (p = 1e-14), and from 61.9% to 100% on ModelNet40/PointMLP (p = 1e-4).

We will add this WSM analysis to the appendix.

AuthorsOpenReview ↗

Response to Reviewer xVWW Part 2

Q3 / W3: What counts and rounding underlie the point-cloud results?

The submitted point-cloud entries were rounded to three significant digits, turning FGSM's ASR of 93/356 = 26.12% into 26.10% and MI-FGSM's ASR of 276/356 = 77.53% into 77.50%. We will correct all affected entries to percentages computed from their k/n ratios and rounded to two decimal places. Every single-run point-cloud entry is computed from a ratio k/n of integer counts, with shared denominators n = 356, 330, 492, and 500 for ModelNet40/PointNet, ModelNet40/PointMLP, ShapeNetPart/PointNet, and ShapeNetPart/PointMLP, respectively. The other single-run entries on ModelNet40/PointNet include NI-FGSM at 129/356 = 36.24%, AWT at 43/356 = 12.08%, and Morpheus at 313/356 = 87.92%.

Q3 / W4: Do the gains persist across independently trained FL trajectories?

Morpheus remains the most effective attack on three independently retrained FL trajectories per modality. We retrained the models with different random seeds, changing the initialization, client partition, and per-round client sampling, and attacked them at the same leakage and deployment stages.

On images (RAF-DB/ResNet18), Morpheus leads the strongest baseline, Proteus, by 2.50 to 5.92 percentage points on the three new trajectories.

AttackTrajectory (Paper)Trajectory (New 1)Trajectory (New 2)Trajectory (New 3)
MI-FGSM55.9154.9352.2247.61
PGD50.5445.0746.3939.44
RaPA with transformations61.2959.1555.8345.35
Proteus with transformations77.9674.0873.0669.58
Morpheus (Ours)82.8080.0075.5672.68

On point clouds (ModelNet40/PointNet), Morpheus leads the strongest baseline, MI-FGSM, by 8.31 to 12.78 percentage points on the three new trajectories.

AttackTrajectory (Paper)Trajectory (New 1)Trajectory (New 2)Trajectory (New 3)
FGSM26.1224.0431.6726.63
PGD61.8057.2757.7856.66
MI-FGSM77.5375.6779.4476.77
NI-FGSM36.2430.5638.3332.58
RaPA75.9368.9475.7472.52
Morpheus (Ours)87.9283.9892.2287.25

On audio (UrbanSound8K/CBAM-DenseNet121), Morpheus leads the strongest baseline, RaPA, by 14.50 to 23.10 percentage points on the three new trajectories.

AttackTrajectory (Paper)Trajectory (New 1)Trajectory (New 2)Trajectory (New 3)
FGSM40.5833.6728.9535.00
PGD65.0863.6754.9358.10
MI-FGSM67.6966.1757.9060.69
NI-FGSM60.3457.3348.6752.24
RaPA74.2778.5071.0566.38
Morpheus (Ours)91.1293.0091.8689.48

Thank you for the detailed response and additional experiments. The matched-budget comparisons and independent training runs address much of my concern about the empirical results. The additional WSM analysis is helpful, and the rounding explanation resolves the point-cloud discrepancy. I am raising my rating from 22 to 33.

Please qualify the image-performance claims, specify the budget used for Proteus, and acknowledge that multi-rate sampling offers no gain over the fixed mean rate on TinyImageNet/ResNet18.

Please also explain why the RAF-DB/ResNet18 WSM correlations differ between the two analyses (0.710.71 versus 1.001.00), including the score definition and comparison sets. These are remaining clarifications and do not change my updated recommendation.

We thank Reviewer xVWW for the encouraging updated assessment and for these remaining comments. We address them below.

On images, Morpheus leads Proteus with transformations on both RAF-DB pairs (82.80% against 77.96% and 56.28% against 54.27%) and trails it on both TinyImageNet pairs (87.20% against 88.15% and 85.71% against 87.01%). In that comparison, Proteus uses approximately 800 gradient evaluations per AE with its original setting of 20 transformed copies per phase, and Morpheus uses 640 with N=32.

At a matched budget of 100 gradient evaluations per AE, log-uniform multi-rate sampling improves ASR by 2.15 points over a fixed rate equal to its mean (0.0195) on RAF-DB/ResNet18 (80.91% against 78.76%). On TinyImageNet/ResNet18, transfer already saturates at a high ASR, so the drop rate has little effect. At this budget, every rate setting we tested yields between 86.26% and 87.68%.

The two RAF-DB/ResNet18 WSM correlations use the same score, the mean raw WSM (Eq. 3), on different comparison sets. The value 0.71 is computed over all seven attacks in the comparison, and the value 1.00 over the five attack families in Fig. 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.