- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper proposes Multi-Round SPEAR++, an exact input-reconstruction attack in federated learning. Building on SPEAR++, the method exploits the sparsity of gradients induced by ReLU activations in fully connected layers to recover training inputs through a low-rank factorization of model updates. The main extension consists of jointly exploiting updates from multiple communication rounds to improve reconstruction success on image datasets.
Critical strengths
-
The paper demonstrates how an honest-but-curious attacker can extend the SPEAR++ framework to exploit model updates across multiple communication rounds, improving reconstruction success.
-
The paper focuses on the FedAvg setting, including multiple local optimization steps, going beyond the simplified single-step FedSGD setting commonly considered in gradient inversion literature.
-
Proposition 1 provides a novel theoretical characterization of the strict local minimizers of the optimization problem in Equation 11, offering insights into the conditions under which input reconstruction can succeed.
Critical weaknesses
-
The paper builds heavily on SPEAR and SPEAR++, reusing essentially the same procedure for candidate selection and recovery of the matrix . The main algorithmic contribution consists of normalizing and stacking model updates from multiple communication rounds into a single matrix, with minor adaptations to the candidate-filtering procedure. This increases the number of available constraints and helps satisfy the rank condition. Consequently, as shown in Figure 3, a major benefit of Multi-Round SPEAR++ is the reduction of rank-related failures F1. Overall, the algorithmic novelty compared to SPEAR++ appears limited.
-
The proposed method does not overcome several important limitations of SPEAR and SPEAR++. In particular, it requires the input matrix to have full column rank, which implies that the input dimension must be at least as large as the number of reconstructed samples . This restricts its applicability to low-dimensional data, such as tabular datasets, when exceeds the input dimension. Moreover, the approach relies on the exact sparsity induced by ReLU activations, and its effectiveness with smooth activation functions, such as GELU or SiLU, and more complex models that do not have a first FC layer remains unexplored. Similarly, the paper does not evaluate robustness to noisy or differentially private model updates.
-
The experimental results do not provide a sufficiently comprehensive picture of the advantages of Multi-Round SPEAR++ over SPEAR++. While the experiments on CIFAR-10 demonstrate some improvements in reconstruction success, the results on MNIST reported in the Appendix are less consistent, with SPEAR++ outperforming the proposed method in some settings. Furthermore, the paper does not provide a detailed analysis of the additional computational cost introduced by jointly processing multiple communication rounds. Consequently, it remains unclear under which conditions the increased computational burden is justified by improvements in reconstruction performance. This concern is particularly relevant for higher-resolution image datasets, such as ImageNet, where stacking and factorizing multiple large gradient matrices could introduce substantial computational and memory overhead.
-
The experimental evaluation does not fully isolate the benefits of the proposed joint multi-round reconstruction approach. Although jointly combining updates is a contribution of the paper, nothing prevents an attacker from applying SPEAR++ independently to each observed round and then combining the resulting reconstructions. Such a baseline would ensure that both attacks have access to the same information and would provide a fairer comparison, allowing the authors to quantify the benefits of jointly processing multiple rounds rather than simply observing additional model updates. Furthermore, this comparison would help assess the computational efficiency of Multi-Round SPEAR++ relative to running SPEAR++ independently across multiple rounds, clarifying whether the proposed approach offers advantages in terms of both reconstruction performance and computational cost
Questions for authors
-
Could the authors elaborate on why Multi-Round SPEAR++ does not consistently benefit from additional communication rounds on MNIST, particularly for smaller local dataset sizes? Intuitively, observing additional rounds should provide more constraints on the shared input factorization that the underlying dictionary-learning problem aims to recover, potentially facilitating the reconstruction. It would be helpful to include an analysis similar to Figure 3 for MNIST, distinguishing the different types of reconstruction failures to better understand the reasons behind this behavior.
-
Could the authors provide a more detailed analysis of the computational cost of the attack, particularly how runtime and memory requirements scale with the input dimension, number of neurons, and the number of observed communication rounds?
-
Could the authors evaluate Multi-Round SPEAR++ against a baseline that independently applies SPEAR++ to the same number of observed communication rounds and then combines the resulting reconstructions? This would help isolate the benefits of the proposed joint reconstruction strategy and provide a clearer comparison of reconstruction performance and computational efficiency.
5 comments
Analysis of Why Multi-Round SPEAR++ Can Underperform SPEAR++ on MNIST (W3, Q1)
Thank you for the suggestion. We conducted an additional failure analysis on MNIST to identify why Multi-Round SPEAR++ does not consistently outperform SPEAR++, as shown in Figure 6 of Appendix F.2. Figure 6 shows that the reduced reconstruction performance in the evaluated setting is primarily associated with completeness failures, followed by candidate-search and dual-margin failures, rather than failures of the rank condition.
These results demonstrate that satisfying the rank condition does not guarantee full reconstruction, which also requires successful candidate search, selection of the complete set of true column directions, and recovery of their signed scales. This explains why additional communication rounds can improve the rank condition without necessarily improving the full recovery rate on MNIST.
Analysis of Computational Overhead and Scalability (W3, Q2)
Thank you for the suggestion. We additionally evaluated reconstruction time and peak GPU memory across different numbers of observed rounds, local data sizes, first-layer widths, and input dimensions, as shown in Table 5 of Appendix F.2. Peak GPU memory increases approximately linearly with , , and , while remaining nearly constant with . Reconstruction time for successful runs generally increases with and but does not exhibit a consistent scaling trend with or . These results clarify the trade-off between computational cost and reconstruction capability, showing that Multi-Round SPEAR++ incurs additional overhead but achieves full recovery in settings where SPEAR++ fails.
Comparison with Independent Reconstruction Across Multiple Rounds (W4, Q3)
Thank you for the suggestion. We confirmed that our joint multi-round reconstruction formulation outperforms independently reconstructing and combining inputs across the same communication rounds, as shown in the newly added Figure 3 in Section 5.2 of the revised manuscript. We applied SPEAR++ independently to each of the first 10 rounds and combined the reconstructed inputs for each client, ensuring that both methods use the same observed updates. These results demonstrate the effectiveness of our joint reconstruction formulation beyond simply observing multiple communication rounds. The computational overhead of Multi-Round SPEAR++ is reported separately in Table 5 of Appendix F.2.
Thank you again for your review. As less than 36 hours remain before the discussion ends, we would appreciate it if you could clarify whether our responses have addressed your concerns and, if not, which concerns remain.
We thank the reviewer for the thoughtful feedback, which helped us substantially strengthen the paper. Our central contribution is a formulation that enables joint exact reconstruction from model updates across multiple communication rounds. We complement this formulation with an algorithm derived from it, a novel theoretical characterization of reconstruction conditions, and additional experiments that isolate the benefits of joint reconstruction. We hope these revisions address your concerns and welcome further discussion.
Novelty and Effectiveness of Joint Reconstruction (W1, W4, Q3)
We would like to clarify that our main contributions are a new multi-round reconstruction formulation and an algorithm derived from it, a novel theoretical characterization of reconstruction conditions, and substantial improvements in exact reconstruction, as summarized in the Introduction.
Our formulation enables joint reconstruction of a common input matrix from model updates observed across multiple rounds. We derive Multi-Round SPEAR++ by adapting SPEAR++ with per-round normalization, block-wise sparsity filtering, and ReLU consistency checks across rounds. To further validate the benefit of our joint reconstruction formulation, as suggested in W4 and Q3, we added a comparison with independently applying SPEAR++ to each observed round and combining the reconstructed inputs. Figure 3 in Section 5.2 of the revised manuscript shows that Multi-Round SPEAR++ outperforms this baseline, demonstrating the effectiveness of our joint reconstruction formulation beyond simply observing multiple rounds.
Furthermore, Proposition 1 provides necessary and sufficient conditions for strict local minimizers of the ℓ1 objective used by both methods. Figure 4 shows different dual-margin failure rates between concatenating rounds and increasing the number of first-layer neurons, even with the same number of rows. These results demonstrate that the benefits of multi-round reconstruction cannot be explained solely by reducing rank-condition failures.
Clarification of the Full-Column-Rank Assumptions on the Input and Gradient Matrices (W2)
We agree that our method inherits the full-column-rank assumption on the input matrix X from SPEAR++, requiring n≥B. Consequently, our method remains limited for low-dimensional data when the number of reconstructed inputs exceeds the input dimension.
However, Multi-Round SPEAR++ relaxes the corresponding rank assumption on the backpropagated-gradient matrix by exploiting multiple communication rounds. As established in Section 3.1, the concatenated gradient matrix can have full column rank even when the gradient matrices from individual rounds are rank deficient. Thus, our formulation relaxes the gradient-side rank requirement while retaining the input-side limitation.
Clarification of Applicability to Other Architectures and Activation Functions (W2)
Thank you for the suggestion. We confirmed that Multi-Round SPEAR++ can reconstruct intermediate features in CNN-based models with a frozen backbone, as shown in Table 3 of Appendix F.2. Using a frozen ResNet-18 backbone and a fully connected ReLU head, Multi-Round SPEAR++ achieves nearly perfect cosine similarity between reconstructed and original features, substantially outperforming SPEAR++. The frozen backbone preserves intermediate features across rounds, allowing our common-input formulation to apply. The recovered intermediate features could be used to approximately reconstruct the original input images through feature inversion, as demonstrated in the original SPEAR paper.
Direct application to convolutional first layers remains challenging because their weight updates factorize over spatial patches rather than individual input images. For GELU and SiLU, the multi-round factorization remains valid for fully connected layers under SGD. However, our reconstruction algorithm cannot be directly applied because these activations do not generally induce the exact sparsity required for reconstruction. Extending the reconstruction algorithm to these activations remains a promising future direction.
Robustness to Clipping and Gaussian Noise (W2)
Thank you for the suggestion. We conducted additional experiments with full-batch DP-GD using per-sample gradient clipping and Gaussian noise, as shown in Figure 14 of Appendix F.2. This result shows that clipping alone does not prevent full reconstruction, and Multi-Round SPEAR++ substantially outperforms SPEAR++ under very small Gaussian noise. In contrast, neither method achieves full recovery when the noise is calibrated to provide a meaningful DP guarantee. Table 4 reports the corresponding privacy budgets.