PAFS: Perturbation-Aware Firing Scaffolds for Robust Event-Based SNNs

Ning Liu; Jun Sun; Hao Guo; Sen Shen; Xiangfei Yang; Jie Yang; Lan Wu; Shangke Lyu

cs.NEOct 23, 2026

PAFS: Perturbation-Aware Firing Scaffolds for Robust Event-Based SNNs

Authors: Ning Liu, Jun Sun, Hao Guo, Sen Shen, Xiangfei Yang, Jie Yang, Lan Wu, Shangke Lyu,

Organizations: The University of Western Australia, Australia · Nanjing University, China · The Chinese University of Hong Kong, Hong Kong SAR, China · Hangzhou City University, China ·

Abstract

Event-based spiking neural networks (SNNs) encode information through temporal firing dynamics, but the same stateful computation makes their predictions sensitive to event loss, timestamp jitter, and speed variation. These perturbations alter evidence accumulation, yet late-layer firing trajectories can retain error-diagnostic structure that is absent from the final logits alone. We introduce Perturbation-Aware Firing Scaffolds (PAFS), a labeled post-training calibration method that exploits this structure without updating the pretrained backbone. PAFS associates standardized late-layer leaky integrate-and-fire trajectories with the class-probability residuals of an Adapter anchor, then applies singular value decomposition to extract compact residual-aligned channel directions. At inference, firing trajectories from the same backbone pass are projected onto this fixed low-rank scaffold and combined with anchor statistics to predict a residual logit correction. Across four matched dataset-backbone settings spanning three event datasets and two spiking Transformer architectures, PAFS improves corrupted-average and worst-condition accuracy over the corresponding Adapter anchors, with maximum gains of 3.06 and 3.65 percentage points. In the DailyAction-DVS/Spikingformer setting, the added trainable parameters amount to approximately 0.34% of the frozen backbone.

Figures & tables

Figure 1: Overview of PAFS. The SNN backbone remains frozen throughout. Stage 1 learns a lightweight Adapter anchor under multiple temporal domains. Stage 2 constructs perturbation-aware firing scaffolds by aligning late-layer LIF trajectories with anchor prediction residuals and extracting dominant error-related firing directions. Stage 3 projects firing responses onto the fixed scaffolds and learns a residual logit correction for the anchor predictor.

DailyAction-DVS, Spikformer
Method	Clean	Window Drop	Jitter	Speed	Corr. Avg.	Worst
LoRA	97.62 $\pm$ 0.55	86.51 $\pm$ 1.40	91.35 $\pm$ 2.10	87.89 $\pm$ 3.00	88.58 $\pm$ 1.90	86.51 $\pm$ 1.40
Plain MLP	97.52 $\pm$ 0.60	86.16 $\pm$ 1.55	89.62 $\pm$ 2.35	87.89 $\pm$ 3.20	87.89 $\pm$ 2.05	86.16 $\pm$ 1.55
Adapter	97.35 $\pm$ 0.53	86.62 $\pm$ 1.22	90.43 $\pm$ 2.50	86.85 $\pm$ 3.65	87.97 $\pm$ 1.57	86.62 $\pm$ 1.22
PAFS (ours)	97.46 $\pm$ 0.53	87.20 $\pm$ 1.51	90.77 $\pm$ 2.80	88.93 $\pm$ 3.34	88.97 $\pm$ 2.28	87.20 $\pm$ 1.51
DailyAction-DVS, Spikingformer

Table 1: Robustness comparison across datasets and spiking Transformer backbones. Results are mean test accuracy (%)

\pm

standard deviation over three matched seeds; Corr. Avg. averages the three temporal corruptions, and Worst denotes the lowest corrupted-condition accuracy.

Dataset	Representation	Clean	Window Drop	Jitter	Speed	Avg.	Corr. Avg.	Worst
CIFAR10-DVS	PAFS (ours)	57.43 $\pm$ 1.21	53.53 $\pm$ 3.00	49.07 $\pm$ 2.40	37.73 $\pm$ 2.66	49.44 $\pm$ 2.63	46.78 $\pm$ 2.63	37.73 $\pm$ 2.66
	PCA scaffold	56.93 $\pm$ 1.08	52.87 $\pm$ 2.55	47.42 $\pm$ 2.08	36.31 $\pm$ 2.18	48.38 $\pm$ 1.92	45.53 $\pm$ 2.03	36.31 $\pm$ 2.18
	Random scaffold	56.84 $\pm$ 1.17	52.61 $\pm$ 2.72	47.18 $\pm$ 2.22	36.74 $\pm$ 2.36	48.34 $\pm$ 2.06	45.51 $\pm$ 2.14	36.74 $\pm$ 2.36
	Logit-only MLP	56.71 $\pm$ 1.12	52.76 $\pm$ 2.63	47.05 $\pm$ 1.96	36.58 $\pm$ 2.27	48.28 $\pm$ 1.98	45.46 $\pm$ 2.07	36.58 $\pm$ 2.27
	Adapter anchor	57.50 $\pm$ 1.32	53.43 $\pm$ 2.84	48.80 $\pm$ 1.95	37.60 $\pm$ 2.43	49.33 $\pm$ 2.34	46.61 $\pm$ 2.34	37.60 $\pm$ 2.43
DailyAction-DVS	PAFS (ours)	99.65 $\pm$ 0.27	94.69 $\pm$ 2.03	96.54 $\pm$ 1.83	94.46 $\pm$ 2.96	96.34 $\pm$ 2.20	95.23 $\pm$ 2.20	94.46 $\pm$ 2.96

Table 2: Mechanism controls with Spikingformer under a matched Adapter anchor and correction protocol. All variants differ only in the representation used for correction. Results are mean test accuracy (%)

\pm

standard deviation over three matched seeds.

Figure 2: Generalization across Window Drop, Jitter, and Speed severities on DailyAction-DVS with Spikingformer. Results are averaged over three matched random seeds, with shaded regions showing

\pm 1

standard deviation.

Figure 3: Rank sensitivity of PAFS on DailyAction-DVS with Spikformer. The dashed line marks the validation-selected rank

r=4

; error bars show

\pm 1

standard deviation over three matched seeds.

Figure 4: Layer evidence on DailyAction-DVS with Spikformer and Spikingformer. Attention+MLP denotes the final PAFS configuration; error bars show

\pm 1

standard deviation over three matched seeds.

Control	Matched components	Changed factor
LoRA	Frozen predictor, data protocol, and validation selection	Parameter-efficient weight update
Plain MLP	Frozen predictor and approximately matched Adapter budget	Output-space residual MLP without an Adapter anchor
Logit-only	Adapter anchor, correction budget, optimization, and selection	Uses only $\phi(z)$ ; no firing features
PCA / Random	Anchor, LIF sources, rank, correction head, and training	Fixed firing-basis construction rule
Scalar rules	Candidate coordinates, retained budget, correction head, and training	Channel-ranking criterion only

Table S1: Matched conditions and isolated factors for the baseline and control comparisons.

Setting	Value
Input representation	$T=16$ event-count frames
Spatial resolution	$128\times 128$
Adapter bottleneck	$b=12$
Adapter dropout	$0.05$
Optimizer	AdamW
Learning rate	$5\times 10^{-4}$

Table S2: Default PAFS implementation, optimization, and model-selection settings. The calibration request cap applies to unique training samples and may be truncated by training-split availability.

Figure S1: Sensitivity to the training-only calibration request cap on DailyAction-DVS with Spikformer. Requested caps of 128, 256, 512, and 1,024 yield realized pools of 128, 256, 512, and 929 unique samples, respectively. The fixed 0.75 fitting fraction gives scaffold-fitting subsets of 96, 192, 384, and 697 samples. Each point is the mean over three matched seeds, with error bars denoting

\pm 1

standard deviation. Corr. Avg. averages Window Drop, Jitter, and Speed; Worst reports the corruption with the lowest cross-run mean.

	Deployment overhead $\downarrow$			Measured inference cost $\downarrow$		Robustness gain vs. Adapter $\uparrow$
Method	Learned params ( $K$ / % frozen)	FP32 storage (KiB)	Added MACs ( $M$ /sample)	Latency (ms/sample)	Energy (mJ/sample)	$\Delta$ Corr. Avg. (pp)	$\Delta$ Worst (pp)
Adapter	$6.924$ / $0.270\%$	$27.0$	$6.291$	$5.989\pm 0.048$	$819.1\pm 11.1$	$0.00$	$0.00$
LoRA	$36.864$ / $1.435\%$	$144.0$	$37.749$ $(+31.458)$	$6.299\pm 0.160$ $(+5.2\%)$	$854.9\pm 12.7$ $(+4.4\%)$	$-1.12$	$+0.12$
Logit-only MLP	$8.796$ / $0.342\%$	$34.4$	$6.293$ $(+0.002)$	$6.318\pm 0.065$ $(+5.5\%)$	$853.3\pm 3.3$ $(+4.2\%)$	$-0.83$	$+0.24$
PAFS (ours)	$8.796$ / $0.342\%$	$106.4$	$6.326$ $(+0.035)$	$6.626\pm 0.011$ $(+10.6\%)$	$880.1\pm 7.4$ $(+7.5\%)$	$\mathbf{+1.84}$	$\mathbf{+3.23}$

Table S3: Efficiency–robustness trade-off on DailyAction-DVS with Spikingformer. All methods share the same frozen

2.569

M-parameter predictor. Learned parameters and FP32 storage denote method-specific additions; PAFS storage additionally includes

72.0

KiB fixed scaffold and normalization buffers. Added MACs are measured relative to the frozen predictor. Latency and energy are mean

\pm

standard deviation over three matched seeds. Parentheses indicate changes from Adapter. Robustness gains are absolute percentage-point improvements. Here,

K=10^{3}

M=10^{6}

, and KiB denotes

2^{10}

bytes.

Method / scoring rule	Clean	Window Drop	Jitter	Speed	Corr. Avg.	Worst
Adapter	97.35 $\pm$ 0.53	86.62 $\pm$ 1.22	90.43 $\pm$ 2.50	86.85 $\pm$ 3.65	87.97 $\pm$ 1.57	86.62 $\pm$ 1.22
PAFS (Ours)	97.46 $\pm$ 0.53	87.20 $\pm$ 1.51	90.77 $\pm$ 2.80	88.93 $\pm$ 3.34	88.97 $\pm$ 2.28	87.20 $\pm$ 1.51
Random ranking	97.18 $\pm$ 0.61	85.74 $\pm$ 1.68	89.88 $\pm$ 2.17	85.92 $\pm$ 3.08	87.18 $\pm$ 1.46	85.74 $\pm$ 1.68
Activation frequency	97.52 $\pm$ 0.48	86.48 $\pm$ 1.45	90.61 $\pm$ 1.92	86.34 $\pm$ 2.84	87.81 $\pm$ 1.30	86.34 $\pm$ 2.84
Activation variance	97.58 $\pm$ 0.46	86.71 $\pm$ 1.56	89.92 $\pm$ 2.06	86.10 $\pm$ 2.73	87.58 $\pm$ 1.28	86.10 $\pm$ 2.73
Temporal variance	97.31 $\pm$ 0.55	87.36 $\pm$ 1.41	89.85 $\pm$ 1.80	86.20 $\pm$ 2.69	87.80 $\pm$ 1.21	86.20 $\pm$ 2.69

Table S4: Controlled comparison of PAFS and scalar channel-selection rules on DailyAction-DVS with Spikformer under the default rank-4-per-layer protocol. PAFS and all scalar controls use the same retained-feature budget and the same 1,872-parameter correction head. Only the rule used to construct the retained firing representation changes. Results are mean test accuracy (%)

\pm

standard deviation over three matched model seeds. Corr. Avg. averages Window Drop, Jitter, and Speed, while Worst denotes the corruption with the lowest cross-run mean accuracy.

Spiking Neural Networks

Spiking neural networks process information through discrete, time-dependent neural events, aiming for efficient computation and biologically inspired learning.

See all 351 papers

Spiking Layer-Adaptive Magnitude-based Pruning
Mar 16, 2026Junqiao Wang, Zhehang Ye, Yuqi OuyangSpiking Neural Networks Neural Network Pruning
Sichuan University-Pittsburgh Institute, Sichuan University, Chengdu, 610207, China · College of Computer Science, Sichuan University, Chengdu, 610065, China
SAGE: Surrogate-gradient Adaptation via Attention-Guided Entropy for Spiking Transformers
Aug 13, 2026Kiran Nair, Rodrigue Rizk, KC SantoshSpiking Neural Networks Neural Network Optimization
USD Artificial Intelligence Research Lab Department of Computer Science University of South Dakota Vermillion, SD 57069, USA
SICAF: Time-Varying Focus Bottleneck for Self-Supervised Event-Based Optical Flow with Spiking Neural Network
Oct 21, 2026Shuangming Yang, Qing He, Shangqi Guo +1Event-Based Optical Flow Spiking Neural Networks
Tianjin University, Tianjin, China · Tsinghua University, Beijing, China · Xi’an Jiaotong University, Xi’an, China

Mar 16, 2026J/K move · Enter open · S save

Spiking Layer-Adaptive Magnitude-based Pruning

Junqiao Wang, Zhehang Ye, Yuqi Ouyang

Sichuan University-Pittsburgh Institute, Sichuan University, Chengdu, 610207, China · College of Computer Science, Sichuan University, Chengdu, 610065, China

Spiking Neural Networks Neural Network Pruning Efficient Neural Network Inference

Open paper AskPDF arXiv

Spiking Neural Networks (SNNs) provide energy-efficient computation but their deployment is constrained by dense connectivity and high spiking operation costs. Existing magnitude-based pruning strategies, when naively applied to SNNs, fail to account for temporal accumulation, non-uniform timestep contributions, and membrane stability, often leading to severe performance degradation. This paper proposes Spiking Layer-Adaptive Magnitude-based Pruning (SLAMP), a theory-guided pruning framework that generalizes layer-adaptive magnitude pruning to temporal SNNs by explicitly controlling worst-case output distortion across layers and timesteps. SLAMP formulates sparsity allocation as a temporal distortion-constrained optimization problem, yielding time-aware layer importance scores that reduce to conventional layer-adaptive pruning in single-timestep limit. An efficient two-stage procedure is derived, combining temporal score estimation, global sparsity allocation, and magnitude pruning with retraining for stability recovery. Experiments on CIFAR10, CIFAR100, and the event-based CIFAR10-DVS datasets demonstrate that SLAMP achieves substantial connectivity and spiking operation reductions while preserving accuracy, enabling efficient and deployable SNN inference.

Abstract

Spiking Neural Networks

Explore similar work

Spiking Layer-Adaptive Magnitude-based Pruning

SAGE: Surrogate-gradient Adaptation via Attention-Guided Entropy for Spiking Transformers

SICAF: Time-Varying Focus Bottleneck for Self-Supervised Event-Based Optical Flow with Spiking Neural Network