SSEL: spike-based structural entropic learning for spiking graph neural networks

Shuangming Yang, Yuzhu Wu, Badong ChenFrontiers in Neuroscience 2025 · DOI

TL;DR

This paper introduces SSEL, a novel framework for Spiking Graph Neural Networks (SGNNs) that enhances robustness against adversarial attacks and significantly improves energy efficiency. By integrating structural entropy theory, SSEL learns to prune noisy graph connections and gate information flow, achieving state-of-the-art performance on both clean and perturbed graphs while drastically reducing energy consumption.

What Problem Does This Paper Address?

Graph Neural Networks (GNNs) are powerful tools for analyzing graph-structured data, but they are vulnerable to adversarial attacks that subtly alter graph connections, leading to performance degradation. Spiking Neural Networks (SNNs), on the other hand, offer remarkable energy efficiency due to their event-driven nature, making them ideal for neuromorphic computing. However, extending SNNs to graphs (Spiking Graph Neural Networks or SGNNs) introduces a new challenge: existing SGNNs are also fragile to these adversarial topology perturbations. This paper tackles the critical gap of achieving both robustness against graph attacks and high energy efficiency in SGNNs, a combination that has been elusive.

Methods

The core innovation of SSEL lies in its integration of structural entropy theory into SGNN learning. This theory helps quantify the inherent structure and community organization within a graph. SSEL employs two key mechanisms:

  1. Structural Entropy-Guided Topology Refinement: By minimizing structural entropy, SSEL learns an "ideal" graph structure that prunes noisy or adversarial edges while preserving connections crucial for community coherence. This is achieved through a differentiable loss function that guides the learning of an adaptive adjacency matrix.
  2. Entropy-Driven Topological Gating: This mechanism uses the refined graph topology to control the flow of information in the SGNN. Only spikes traveling along the "entropy-optimized" edges are allowed to propagate, effectively blocking adversarial pathways and ensuring that computations are focused on meaningful connections.

This dual approach synergizes structural sparsity (from the pruned graph) and event-driven sparsity (inherent to SNNs) to achieve both robustness and efficiency.

Datasets

The paper evaluates SSEL on two benchmark citation network datasets:

  • Cora: Contains 2,708 publications and 5,429 citation links.
  • Citeseer: Contains 3,327 publications and 4,732 citation links.

These datasets are standard for evaluating graph-based learning models.

Key Insights and Ideas

  • Structural Entropy as a Robustness Lever: The paper demonstrates that minimizing structural entropy is not just about finding communities, but a principled way to identify and remove noisy or adversarial connections in graphs, leading to inherent robustness.
  • Co-design for Dual Sparsity: The true power of SSEL comes from its co-design approach, where the learned graph structure directly influences the spiking dynamics. This creates a synergistic effect where structural sparsity enhances event-driven sparsity and vice-versa.
  • Topological Gating is Crucial: The entropy-driven gating mechanism acts as a critical filter, ensuring that even if a graph has some noisy edges, the SGNN only processes information through the robust, entropy-optimized pathways.
  • Bridging the Gap: SSEL successfully bridges the gap between the energy efficiency of SNNs and the robustness required for real-world graph applications, which often face adversarial perturbations.

Key Results

  • Exceptional Robustness: Under a 0.1 salt-and-pepper noise attack, SSEL achieved 64.58% accuracy, significantly outperforming the baseline SGNN (w.o. SSEL) which dropped to 30.14%.
  • State-of-the-Art Accuracy on Clean Graphs: SSEL achieved 85.31% accuracy on the Cora dataset, outperforming various GNN and SGNN baselines.
  • Ultra-Low Energy Consumption: SSEL achieved a 97.28% reduction in energy consumption compared to conventional GNNs, while maintaining high accuracy.
  • Significant Improvement over Baselines: On Citeseer, SSEL showed a 7.59% absolute improvement over Spiking GCN.
  • Dataset Dependency: The paper notes that the benefits of structural entropy might be dataset-dependent, with more pronounced advantages on graphs with inherently noisier structures.

Limitations and Future Work

The authors acknowledge the following limitations and suggest future research directions:

  • Scalability: The computational overhead of entropy minimization might impact performance on very large-scale graphs. Future work will focus on improving scalability through adaptive thresholds and efficient algorithms.
  • Static Graphs: The current framework is designed for static graphs. Future research will explore its application to dynamically evolving topologies.
  • Distributed Training: Developing distributed training strategies for neuromorphic hardware is also a future research avenue.
  • Exploring Other Entropy Measures: Investigating other entropy measures for specific graph types is suggested.

Summary updated

Abstract

Spiking Neural Networks (SNNs) offer transformative, event-driven neuromorphic computing with unparalleled energy efficiency, representing a third-generation AI paradigm. Extending this paradigm to graph-structured data via Spiking Graph Neural Networks (SGNNs) promises energy-efficient graph cognition, yet existing SGNN architectures exhibit critical fragility under adversarial topology perturbations. To address this challenge, this study presents the Spike-based Structural Entropy Learning framework (SSEL), which introduces structural entropy theory into the learning objectives of SGNNs. The core innovation establishes structural entropy-guided topology refinement: By minimizing structural entropy, we derive a sparse topological graph that intrinsically prunes noisy edges while preserving critical low-entropy connections. To further enforce robustness, we develop an entropy-driven topological gating mechanism that restricts spiking message propagation exclusively to entropy-optimized edges, systematically eliminating adversarial pathways. Crucially, this co-design strategy synergizes two sparsity sources: Structural sparsity from the entropy-minimized graph topology and Event-driven sparsity from spike-based computation. This dual mechanism not only ensures exceptional robustness (64.58% accuracy vs. 30.14% baseline under 0.1 salt-and-pepper noise) but also enables ultra-low energy consumption, achieving 97.28% reduction compared to conventional GNNs while maintaining state-of-the-art accuracy (85.31% on Cora). This work demonstrates that the principled minimization of structural entropy is a powerful strategy for enhancing the robustness of Spiking Graph Neural Networks. The SSEL framework successfully mitigates the impact of adversarial topological perturbations while capitalizing on the energy-efficient nature of spike-based computation, which underscore the significant potential of combining information-theoretic graph principles with neuromorphic computing paradigms.
BibTeX
@inproceedings{yang2025ssel,
  author = {Shuangming Yang and Yuzhu Wu and Badong Chen},
  title = {{SSEL: spike-based structural entropic learning for spiking graph neural networks}},
  booktitle = {Frontiers in Neuroscience},
  year = {2025},
  doi = {10.3389/fnins.2025.1687815},
}
Figure