acceptodds
Withdrawn from ICLR 2027

Spikes Leak: Analytical Gradient Inversion in Spiking Neural Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spiking neural networks (SNNs) are promising candidates for federated edge learning, as their event-driven computation makes them well suited to energy-constrained devices. However, their exposure to gradient inversion attacks is less understood than in conventional artificial neural networks (ANNs). Recent studies, largely based on adapting ANN-oriented attacks to the spiking domain, have suggested that SNN-specific mechanisms may offer inherent privacy advantages. We challenge this claim by examining the algebraic structure of SNN gradients. In fully connected spiking layers, gradients factorize by construction into a temporal gradient matrix and a binary presynaptic spike matrix. This binary factor turns reconstruction into a finite constrained combinatorial problem, creating an exploitable leakage channel. Based on this observation, we propose Spike-Leak, a two-stage attack. It first recovers binary candidate spike patterns from gradient row-space constraints, then reconstructs private input sequences by assigning candidates to examples and timesteps using the known SNN dynamics and downstream consistency. Across four event-based vision datasets, Spike-Leak achieves exact reconstruction across diverse configurations spanning multiple temporal horizons and batch sizes, including a setting with \(B=32\). These results show that SNNs expose a structured gradient-leakage channel, and that binary spike processing, surrogate-gradient training, and temporal aggregation do not, by themselves, prevent exact gradient-based reconstruction.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Spiking Neural Networks, Gradient Inversion Attacks, Federated Learning, Privacy in Machine Learning

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

This paper proposes SpikeLeak, an analytical gradient inversion attack on SNNs. It combines binary candidate recovery from gradient row-space constraints with sequence reconstruction guided by neuronal dynamics and cross-layer consistency. Theoretical analysis establishes rank conditions ensuring coverage of all true spike patterns, while experiments on four event-based vision datasets demonstrate exact reconstruction across multiple configurations.

Critical strengths
  1. Theoretical characterization of gradient leakage. The paper characterizes gradient leakage in fully connected SNNs and establishes rank conditions ensuring that the candidate set contains all true spike patterns.

  2. The two-stage framework combines binary candidate recovery with neuronal dynamics and cross-layer consistency to resolve temporal ordering and sample assignments.

  3. Experiments on four event-based vision datasets and demonstrate exact reconstruction at batch sizes up to 32 on N-MNIST. The sparsity analysis further distinguishes candidate coverage from practical recoverability.

Critical weaknesses

Limited validation on trained models. The main experiments use randomly initialized, fully connected SNNs. It remains unclear whether the rank conditions and reconstruction performance persist after training, limiting the empirical support for applicability to trained models.

Insufficient characterization of uniqueness. The rank analysis guarantees candidate coverage, while exact recovery additionally depends on uniqueness. The paper does not provide a verifiable criterion for determining when the reconstruction constraints uniquely identify the input batch.

High computational cost. Reconstruction takes approximately 20 hours at B=32, T=8, and search can exhaust its budget even with complete candidate coverage. These results leave the computational feasibility of recovery beyond the demonstrated settings uncertain.

Questions for authors
  1. How do gradient rank, candidate coverage, and reconstruction success change across trained checkpoints of the same architecture?

  2. Can the authors provide theoretical or empirical evidence for reconstruction uniqueness up to sample permutation? Are alternative feasible input batches investigated?

  3. For configurations with complete candidate coverage but failed reconstruction, how does reconstruction success vary with the Stage II search budget? This analysis would help determine whether the search budget is the primary bottleneck.

Reviewer_1BErOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SpikeLeak, an analytical gradient inversion attack for spiking neural networks. The method makes use of the gradient structure in SNNs, where continuous backward signals are multiplied by binary presynaptic spikes. Based on this observation, the authors propose a two-stage reconstruction framework. It first identifies candidate spike patterns through row-space constraints and MILP, and then reconstructs temporal inputs using neuronal dynamics and consistency checks. Experiments on four event-based datasets report exact reconstruction in all 57 completed configurations.

Critical strengths
  1. The motivation is clear and relevant. Whether SNNs are inherently resistant to gradient leakage is an interesting security question that deserves further investigation.

  2. The proposed reconstruction framework is reasonably designed. The use of gradient structure to recover spike patterns, followed by temporal reconstruction based on SNN dynamics, is interesting. The theoretical analysis also provides some useful insights into gradient leakage in SNNs.

  3. The experiments show exact reconstruction across different event-based datasets, batch sizes, and temporal settings. The additional analysis of spike sparsity and reconstruction failures is also helpful for understanding the behavior and limitations of the method.

Critical weaknesses
  1. The practical applicability of the attack is quite limited. Most experiments use randomly initialized fully connected SNNs rather than trained convolutional or transformer-based architectures. The attack also assumes full access to gradients, known initial neuronal states, and the absence of batch-coupled operations. These assumptions may not hold in realistic federated learning scenarios, making it difficult to assess the practical significance of the attack.

  2. The scalability of the method is a major concern. The reconstruction process involves expensive combinatorial search and backtracking. For example, the B=32, T=8 setting requires around 19.9 hours, and the method fails in some larger temporal settings due to incomplete candidate coverage or search budget limitations. It is unclear whether the approach can scale to more realistic network sizes and longer sequences.

  3. The comparisons with existing gradient inversion methods are insufficient. Although the paper discusses DLG, iDLG, SPEAR, and DAGER, direct experimental comparisons under the same settings are missing. Also, the gradient factorization used in the analysis is essentially a consequence of the standard chain rule. The authors should better clarify what is methodologically new and how SpikeLeak differs from or improves upon existing inversion approaches.

  4. The theoretical guarantees are somewhat limited. The analysis mainly establishes candidate coverage under certain rank conditions, but does not guarantee unique reconstruction or efficient recovery. In addition, the reported 57/57 success rate only considers completed configurations, while some more challenging settings are not successfully solved. More discussion of reconstruction uniqueness and overall success rates would make the conclusions more convincing.

  5. The experimental settings are still relatively idealized. The paper does not investigate noisy, quantized, clipped, or aggregated gradients after multiple local training steps. These are common conditions in practical federated learning and could significantly affect reconstruction performance. Evaluating at least some of these settings would help establish the robustness of the proposed attack.

Questions for authors

The authors should address these weaknesses carefully.

Reviewer_6FyTOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SpikeLeak, a two-stage analytical gradient inversion attack for spiking neural networks (SNNs). It first recovers binary spike candidates from gradient row-space constraints and then reconstructs input sequences using spatio-temporal consistency. Experiments on four event-based vision datasets demonstrate exact reconstruction across 57 completed configurations, including batch sizes up to 32.

Critical strengths
  1. The paper exploits the binary structure of SNN gradients to enable exact input reconstruction, providing a useful perspective on privacy leakage in SNNs.
  2. The method achieves exact reconstruction across multiple datasets, batch sizes, and temporal horizons, with additional analyses of spike and gradient sparsity.
Critical weaknesses
  1. Experiments rely on randomly initialized, fully connected SNNs, without evaluating trained models, convolutional architectures, or realistic federated learning updates.
  2. The paper lacks direct comparisons with existing gradient inversion attacks under matched experimental settings. Furthermore, the computational cost increases substantially with batch size, with the largest configuration requiring approximately 20 hours.
Questions for authors
  1. Does SpikeLeak remain effective on trained SNNs or more commonly used architectures such as convolutional SNNs? How would the attack perform when only model updates after multiple local training steps are available?
  2. Could the authors compare SpikeLeak with adapted gradient inversion methods such as DLG[1], iDLG[2], or GRNN[3] under matched experimental settings? Reporting both reconstruction quality and computational cost would help establish the practical advantage of the proposed approach.
  3. Section 3 specifies a Stage I enumeration limit of 3,000 seconds per layer, whereas Table 3 in the appendix reports 900 seconds. Could the authors clarify which configuration was used for the reported experiments and runtimes?
  4. Figure 3 shows that reconstructing a batch of 32 samples with T = 8 requires approximately 20 hours. Furthermore, reconstruction fails at T = 36 due to the search budget limit (Table 5). Could the authors discuss the scalability limitations of SpikeLeak and potential ways to improve its computational efficiency?
  5. The reconstruction algorithm returns the first feasible solution without guaranteeing its uniqueness. Could the authors clarify whether multiple feasible reconstructions can exist and evaluate how frequently such ambiguity occurs in practice?

[1] Zhu, L., Liu, Z., & Han, S. (2019). Deep Leakage from Gradients. NeurIPS 2019.

[2] Zhao, B., Mopuri, K. R., & Bilen, H. (2020). iDLG: Improved Deep Leakage from Gradients. arXiv:2001.02610.

[3] Ren, H., Deng, J., & Xie, X. (2022). GRNN: Generative Regression Neural Network—A Data Leakage Attack for Federated Learning. ACM Transactions on Intelligent Systems and Technology, 13(4).

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

Related papers

Discussion (0)

Sign in to comment.