Raw event-based adversarial attacks for Spiking Neural Networks with configurable latencies

Area of Science:

• Artificial Intelligence
• Computer Vision
• Neuromorphic Computing

Background:

• Spiking Neural Networks (SNNs) and Dynamic Vision Sensors (DVSs) provide energy-efficient computing for edge devices.
• Input latency in SNNs is crucial for energy savings, but security vulnerabilities, especially adversarial attacks, are a growing concern.
• Attacks on raw event streams with configurable latencies are underexplored due to data sparsity and discrete optimization challenges.

Purpose of the Study:

• To propose a novel adversarial attack method targeting raw event streams in SNNs with configurable latencies.
• To address the challenges of sparsity and discrete optimization in raw event stream attacks.
• To enhance the security of energy-efficient edge AI systems.

Main Methods:

• Developed a method to smooth optimization by converting binary spikes to continuous values.
• Introduced an adaptively stochastic strategy for sampling attack latencies.
• Applied regularization to maintain sparsity and ensure adversarial samples resemble raw data while approximating target labels.

Main Results:

• The proposed method significantly outperforms existing approaches in adversarial attack success rates (ASR) across various latencies.
• Experiments on N-MNIST, CIFAR10-DVS, N-Caltech-101, and Gesture-DVS datasets demonstrate consistent superiority.
• Ablation studies confirm the effectiveness of the proposed techniques and the influence of latency on adversarial sample generation.

Conclusions:

• The novel attack method effectively targets raw event streams in SNNs with configurable latencies.
• The approach enhances the security of event-based vision systems against adversarial manipulation.
• This research opens new avenues for understanding and defending against attacks on neuromorphic systems.