acceptodds
Withdrawn from ICLR 2027

Cryptanalytic Extraction of Direct-Coded LIF Spiking Neural Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spiking neural networks (SNNs) are increasingly deployed on edge and neuromorphic platforms, making their parameters and temporal dynamics valuable intellectual property. Cryptanalytic model extraction has focused mainly on conventional neural networks, leaving structural leakage from discrete firing, membrane recurrence, and reset dynamics in LIF-SNNs underexplored. We study cryptanalytic extraction of direct-coded LIF-SNNs through a black-box interface exposing only final logits. We show that LIF dynamics create structured input-space boundaries corresponding to spike-count transitions; crossing an isolated boundary associated with a hidden neuron produces an observable logit jump whose direction is determined by that neuron’s output-weight vector, thereby exposing internal parameter information. Exploiting this observation, our attack performs transition attribution, boundary recovery, and dynamical parameter inversion to recover normalized hidden-layer parameters, the shared leak parameter, and output-layer parameters. The resulting canonical parameterization is equivalent to that of the target model and defines a functionally equivalent model. We further establish sufficient conditions for identifying the canonical parameters. Experiments show complete hidden-neuron recovery, over 99% spike-count fidelity, and near-perfect prediction agreement on natural inputs. This work presents the first systematic study of cryptanalytic extraction for direct-coded LIF-SNNs and demonstrates the need for defenses tailored to SNN model extraction.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
spiking neural networks, model extraction, cryptanalysis, parameter recovery, LIF neurons

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_HBQHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies parameter extraction from single-hidden-layer, direct-coded LIF spiking neural networks using only adaptive input queries and final logits. Its central observation is that constant-drive spike-count transitions form parallel hyperplane families, while isolated crossings produce logit jumps proportional to the corresponding output-weight columns. The proposed attack groups transitions by jump direction, recovers boundary normals and locations, unfolds merged transitions using jump multiplicities, and fits theoretical threshold sequences to recover normalized hidden parameters and a shared leak. Output parameters are subsequently recovered through a linear system. The paper provides conditional identifiability results and evaluates extraction on MNIST and Fashion-MNIST across hidden widths and temporal settings. It reports complete estimate coverage, small normalized-weight and leak errors, above 99% spike-count fidelity, and high prediction agreement, with query costs in the millions.

Critical strengths
  1. A coherent mechanism linking observable logits to hidden temporal dynamics. The paper goes beyond behavioral imitation by exploiting boundary directions, jump multiplicities, and relative boundary spacing to reconstruct canonical parameters. Recovering the shared leak alongside static weights is a consequential aspect of the contribution, and the scale and permutation ambiguities are explicitly addressed.
  2. The experiments assess internal recovery rather than prediction agreement alone. The main study reports normalized-weight error, leak error, and hidden spike-count fidelity across five runs for each hidden width. The transition-unfolding ablation is particularly informative: removing it preserves estimate coverage but substantially degrades transition-index accuracy, leak recovery, and functional fidelity. This supports the importance of accounting for merged transitions.
Critical weaknesses
  1. The central dynamical identifiability result leaves the decisive uniqueness condition unresolved. Theorem 2 assumes unique transition unfolding/alignment and that the selected spacing ratios uniquely determine the leak. The paper explicitly acknowledges that these properties need not hold generally, but does not characterize concrete regimes where they hold or provide a practical certificate for the experimental targets. Thus, the theorem establishes conditional recovery without resolving when the main ambiguity can actually be eliminated. The algorithm additionally relies on selecting among transition assignments, yet its search complexity and behavior under competing leak candidates are insufficiently specified. This limits the theoretical contribution and confidence in recovery beyond the tested configurations.

  2. The validation does not fully support complete canonical recovery or characterize practical robustness. “Recovered” measures estimate coverage without an accuracy threshold, and direct errors are reported for normalized input weights and the leak, but not hidden biases, output weights, or output biases. Functional checks are limited to natural inputs despite the stated equivalence domain being the full query box. Moreover, extraction uses roughly 8.5–96.8 million queries on small, single-hidden-layer models, without runtime or output-precision sensitivity measurements. These limitations are acknowledged, but leave the practical significance of the attack insufficiently established. There is also an evaluation ambiguity: individual configurations in Table 3 report PMR values such as 0.9999, which cannot arise from the stated indicator average over exactly 1,000 test inputs without additional samples or aggregation.

Questions for authors
  1. For the tested horizons and leak values, can the authors establish or empirically certify uniqueness of the aligned threshold pattern over the actual admissible leak interval? How are multiple indistinguishable or near-indistinguishable candidates detected and handled?
  2. Please specify the leak-search interval and resolution, transition-subsequence search, alignment tolerances, candidate-selection loss, and stopping criteria. What are the runtime and query-cost breakdowns for boundary discovery, geometry recovery, and alignment? 3.What exact number of examples and aggregation procedure produce the individual target accuracies and PMRs in Table 3? Please reconcile these values with the stated 1,000-example test set.
  3. Can the authors report errors for all canonical parameter blocks and fidelity on held-out inputs spanning the query domain, including near recovered boundaries? A controlled experiment with reduced logit precision would also help establish whether the demonstrated recovery survives a modest departure from the ideal interface.
Reviewer_7WamOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper combine the Cryptanalytic Model Extraction with direct-coding LIF SNNs. The authors treat LIF-SNNs as an ANN with specific step activation function and then perform Cryptanalytic Model Extraction attack. Experiments on MNIST and FMNIST demonstrate high parameter-recovery accuracy and near-perfect prediction.

Critical strengths

Cryptanalytic model extraction is an interesting research direction as a new type of model attack and AI security.

The paper provides mathematical analysis of parameter identifiability and recovery conditions.

The authors provide a reasonable experimental evaluation over MNIST datasets.

Critical weaknesses

The major weakness of this paper is lack of novelty and significance. This research seems to be an incremental extension of existing cryptanalytic model extraction methods for ANNs. The proposed approaches include adopting direct coding with static inputs and spike-count readout, and then adopt the existing ANN boundary-based parameter recovery techniques. It is therefore difficult to identify what new scientific insight or practically meaningful security vulnerability this work reveals beyond what is already understood from cryptanalytic extraction of ANNs.

Moreover, the direct coding of SNN, as a simplified formulation, avoids the essential temporal characteristics of SNNs. The evalution also limits in static image classification task rather than sequential tasks or dvs input tasks.

It is therefore unclear whether the work reveals any meaningful vulnerability specific to SNNs, rather than simply demonstrating the applicability of an existing ANN extraction framework to another activation function.

Questions for authors
  1. Given existing extraction methods for step activation functions by Qi et al. (2026), what fundamental technical challenges does this work address that cannot be solved by previous approaches?

  2. What is the practical security relevance of requiring millions of queries to extract a tiny single-hidden-layer SNN model? Would the proposed attack remain effective under more realistic neuromorphic hardware conditions, such as quantized weights/outputs event-drive/spike inputs?

Reviewer_6MeWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present a model extraction attack on spiking neural networks. The attack aims to recover LIF neurons using a black-box approach where the attacker has access to the inputs and the outputs' logits. Based on how LIF neurons work, by probing the input, the logits will show a large change when the input has crossed a boundary. They use that information to reconstruct the weights, and the experiments show complete hidden-neuron recovery, with over 99% spike-count fidelity and near-perfect prediction agreement on natural inputs. Finally, the authors establish sufficient conditions for identifying the canonical parameters.

Critical strengths

The first attempt at a model extraction attack on SNNs.

Critical weaknesses

The novelty differentiating this work from previous model extraction attacks on NNs is limited.

Experimentation is limited.

Questions for authors

The paper reads as a similar approach to Carlini et al.'s cryptanalytic extraction. What are the explicit differences? The related work briefly mentions parallel count-boundary families, integer jump multiplicities, and leak-dependent spacing. I would like a direct comparison with the ReLU critical-point setup (as used in Carlini's), stating what is new.

The analysis assumes "a single hidden layer, deterministic soft reset, a shared leak, a continuous bounded query domain, and unquantized final logits." How realistic are these assumptions? Section 7 lists them as limitations, but there is no discussion of how likely they are in deployed systems. Please discuss this.

Only a single hidden layer is considered. Why? This is acknowledged as a limitation, but no experiments are provided. I would say that an ablation on other architectures is required.

The attacker is assumed to know the LIF update rule, including the soft reset. Neurons can use other reset types and decay formulations. What happens if the attacker does not know which one is used? Does the method still work if another reset rule is used?

Only direct coding is studied. Why not other encodings, such as rate coding? What would break in the method under those encodings?

The attack requires the output logits. What happens if the attacker only has the label? The related work cites hard-label extraction attacks, but there is no discussion of whether this method could work in that setting.

What is the real use case of an SNN that is this small? An attacker could train a similar SNN in minutes, and the targets here are trained on only 4,000 samples. What kind of data or model is valuable enough to justify extracting it? The threat model needs a better explanation.

On the experimentation:

MNIST and Fashion-MNIST are used. Why not N-MNIST or another neuromorphic variant? If the method requires static inputs, please state this explicitly and discuss whether neuromorphic data could or could not be used.

The inputs are scaled by 1.5. Why?

"Natural samples are reserved for evaluation". What are natural samples?

In Table 1, the attacker needs more queries for h=32 (46.8M) than for h=64 (19.1M). Intuitively, it should be the other way around. The paper only attributes this to "boundary accessibility," which is not an explanation. The standard deviation for h=64 is also very small (±0.1).

Is there a query budget or cap? Query counts differ a lot across settings. What are the stopping conditions of the search? Algorithm 1 does not say when boundary discovery ends.

What is the accuracy of the extracted model after the weights have been recovered? Only target accuracy, PMR, and logit RMSE are reported.

Only the weight error is reported. Are the hidden and output biases also recovered, and what are their errors?

Defenses are only mentioned in one sentence in Section 7, without evaluation. A discussion is missing, ideally with a small experiment, for example with reduced logit precision or output noise.

Writing

The writing is entangled and packed. Many sentences carry a lot of information, which makes the paper hard to read. One example is "Nonzero pairwise non-proportional output columns are required when jump directions are used for neuron attribution." I suggest rewriting it with a more natural flow and shorter explanations, which would greatly help the reader (especially if not so familiar with the domain).

In the threat model, T is called the "inference horizon," while Section 3.1 defines it as the number of timesteps. Please use one term consistently.

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

Related papers

Discussion (0)

Sign in to comment.