TL;DR
This paper introduces HIRE-SNN, a novel training method for deep Spiking Neural Networks (SNNs) that significantly enhances their robustness against adversarial attacks without increasing training time. By strategically adding crafted noise to input data during training, HIRE-SNN achieves improved accuracy on attacked images while maintaining high performance on clean data and offering substantial gains in energy efficiency and latency compared to existing SNN approaches.
What Problem Does This Paper Address?
Artificial Neural Networks (ANNs) and, more recently, Spiking Neural Networks (SNNs) are increasingly used in safety-critical applications. However, both are vulnerable to adversarial attacks, where small, imperceptible perturbations to input data can cause misclassifications. While SNNs offer significant energy efficiency advantages, especially on neuromorphic hardware, their robustness against these attacks, particularly low-latency SNNs that use direct input encoding, has been underexplored. Existing methods to improve SNN robustness often lead to high inference latency or increased training costs, creating a need for efficient and effective defense mechanisms.
Methods
The authors first conduct an empirical analysis to understand the inherent robustness of low-latency, direct-input SNNs against common gradient-based attacks like FGSM and PGD. Based on these findings, they propose HIRE-SNN, a training algorithm that leverages Spike Timing Dependent Backpropagation (STDB). The core innovation is the introduction of crafted input noise during training. Instead of processing the same image over the entire inference time steps ($T$), HIRE-SNN divides $T$ into smaller periods. Within each period, a slightly perturbed version of the input image is fed to the network, and the weights are updated. This process effectively trains the SNN to be resilient to variations in input data without requiring additional training time, as the noise is introduced within the existing inference steps. They use a linear surrogate gradient approximation to enable backpropagation through the spiking neurons.
Datasets
The paper uses the following standard image classification datasets:
- CIFAR-10
- CIFAR-100
Key Insights and Ideas
- Direct-input SNNs are surprisingly vulnerable: Contrary to some prior assumptions, low-latency SNNs using direct input encoding show lower accuracy than their ANN counterparts when subjected to white-box and black-box adversarial attacks. This highlights a gap in understanding SNN robustness.
- Crafted noise is key: Simply reducing inference time steps in SNNs doesn't automatically confer robustness. The HIRE-SNN approach demonstrates that strategically introducing noise during training is crucial for building resilience.
- No extra training time needed: A significant advantage of HIRE-SNN is that it achieves robustness improvements without increasing the overall training duration, making it practical for real-world deployment.
- LIF neuron non-linearity's complex role: The Leaky Integrate-and-Fire (LIF) neuron's leak parameter, while beneficial in rate-coded SNNs for reducing perturbation distance, has a more complex and sometimes detrimental effect in direct-input SNNs, suggesting that its impact is highly dependent on other model parameters.
- Energy efficiency gains are substantial: HIRE-SNN not only improves robustness but also maintains or enhances the inherent energy efficiency of SNNs, offering significant advantages over both traditional SNNs and ANNs.
Key Results
- Improved Adversarial Accuracy: HIRE-SNN trained models achieve up to 13.7% and 10.1% higher accuracy on FGSM and PGD attack-generated images, respectively, compared to standard direct-input SNNs.
- Negligible Clean Accuracy Drop: The proposed method results in only a minor loss in accuracy on clean images.
- Outperforms Rate-Coded SNNs: HIRE-SNN models show improved or similar performance on attacked images compared to inherently-robust rate-coded SNNs, while offering significantly lower latency (up to 25x lower) and computation energy (up to 4.6x lower).
- Reduced Memory Usage: The HIRE-SNN training approach requires up to 25% less memory due to shorter update intervals.
- Robustness against various attacks: The method shows consistent improvements against both white-box (WB) and black-box (BB) attacks.
Limitations and Future Work
- Gradient Obfuscation: While the authors perform tests to check for gradient obfuscation, they acknowledge that it might still play a role, though not a significant one, in the robustness of HIRE-SNNs. Further research could explore this more deeply.
- Noise Crafting Nuances: The paper highlights the importance of "carefully crafted noise" and shows that simple Gaussian noise is less effective than their proposed method. Understanding the optimal noise characteristics for different architectures and tasks remains an area for exploration.
- Hyperparameter Tuning: The effectiveness of the noise step size ($\epsilon_s$) is demonstrated, but optimal values might vary across different models and datasets.
- Future Work: The authors suggest that their framework can be extended to support various input encoding schemes and image augmentation techniques, potentially further enhancing robustness and performance. They also believe their work is a step towards making deep SNNs practical for safety-critical applications where robustness is paramount.
Summary updated