TL;DR
This paper introduces a novel "spatiotemporal trigger" for backdoor attacks on Spiking Neural Networks (SNNs) trained on event-based data from Dynamic Vision Sensors (DVSs). The attack achieves near-perfect success rates with minimal impact on normal performance, demonstrating a significant new security vulnerability for edge AI systems.
What Problem Does This Paper Address?
Spiking Neural Networks (SNNs) are increasingly used in energy-constrained edge devices for tasks like autonomous driving and privacy-preserving surveillance, thanks to their efficiency and ability to process data from event cameras (DVSs). These cameras, unlike traditional RGB cameras, output asynchronous streams of "events" indicating pixel brightness changes. While SNNs offer advantages, they, like other deep learning models, are susceptible to data poisoning attacks, specifically backdoor attacks. These attacks involve subtly altering training data to create a hidden "trigger" that causes the model to misbehave in a specific, attacker-controlled way during inference, while behaving normally otherwise. Existing backdoor attacks on SNNs are limited, often relying on converting event data into frames (losing temporal information) or requiring full control over the training process. This paper addresses the gap by demonstrating a practical, black-box backdoor attack directly on event data for SNNs, which is more stealthy and robust.
Methods
The core innovation is the spatiotemporal trigger. Instead of relying on image-like patches, this trigger is a sequence of events injected directly into the event stream. Key aspects of the method include:
- Spatiotemporal Nature: The trigger is designed to leverage the inherent temporal dynamics and spatial information of event data, making it harder to detect and remove. It's not just a static pattern but a dynamic sequence.
- Event-Based Injection: The trigger is injected directly into the raw event stream, avoiding the need for event-to-frame conversion, which can lose crucial information and make attacks easier to detect.
- Stealthiness: The trigger is designed to blend seamlessly with the natural event data, making it inconspicuous. The authors explore three different spatial locations for injecting the trigger: upper-left corner, background, and foreground regions of the event scene.
- Robustness: The trigger is designed to withstand common defenses like pruning (removing less important neurons) and event noise filtering, which are often used to mitigate backdoor attacks.
- Black-Box Attack: The attacker does not need knowledge of the SNN's architecture or the specific event representation method used during training. They only need to inject poisoned data into the training set.
Datasets
The proposed attack was evaluated on three standard event-based datasets:
- NMNIST: A neuromorphic version of the MNIST handwritten digit dataset.
- CIFAR10-DVS (C10-DVS): An event-based version of the CIFAR-10 image classification dataset.
- DVS128-Gesture (DVS-G): A dataset for gesture recognition using event cameras.
The paper also mentions that event streams are converted into event frames using two common methods: slicing in time and slicing in the number of events.
Key Insights and Ideas
- Event Data is a Rich Target: The temporal and spatial richness of event data offers new avenues for sophisticated backdoor attacks that are difficult to achieve with static image data.
- Spatiotemporal Triggers are More Stealthy and Robust: By directly manipulating the event stream in a temporally coherent manner, the trigger is less likely to be detected by standard defenses or human inspection compared to frame-based attacks.
- Location Matters (Subtly): While all tested trigger locations were effective, placing the trigger in the background region sometimes yielded slightly better results, possibly because background features are less correlated with task-specific features, making the trigger more independent and harder to remove.
- Minimal Impact on Benign Accuracy: A key finding is that the attack can achieve very high success rates with only a negligible drop in the model's accuracy on clean, untriggered data, making it a truly insidious threat.
Key Results
- High Attack Success Rate (ASR): The proposed spatiotemporal trigger achieved near 100% ASR on various SNN models and datasets, even at a low injection ratio of 0.1 (10% of training data poisoned).
- Minimal Benign Accuracy (BA) Degradation: The BA of the backdoored models remained very close to the baseline accuracy of clean models, indicating normal functionality for non-attacked inputs.
- Robustness to Defenses: The attack demonstrated high resilience to pruning defenses and event noise filtering, with ASR remaining high even after these mitigation attempts.
- Outperforms Prior Work: The spatiotemporal trigger was shown to be more effective and stealthy than previous backdoor attacks on SNNs that relied on frame-based representations.
Limitations and Future Work
- Stealthiness Metric: The paper acknowledges the lack of a definitive quantitative metric for stealthiness in event data, relying on SSIM scores of 2D representations as a proxy.
- Future Work: The authors suggest exploring effective countermeasures to prevent spatiotemporal vulnerabilities from being exploited in future adversarial attacks on SNNs. This implies a need for new defense mechanisms specifically designed for event-based systems.
Summary updated