Spiking Neural Networks (SNNs) paired with event-based sensors offer an energy-efficient, biologically plausible alternative to traditional Artificial Neural Networks (ANNs). Instead of processing continuous-valued pixel arrays frame by frame, SNNs operate on sparse, asynchronous streams of binary spikes. Historically, this discrete, time-dependent processing led to a widespread assumption that SNNs possess an inherent robustness to adversarial perturbations.
However, recent research demonstrates that the temporal dimension of neuromorphic computing is not a shield, but rather a unique and highly sensitive attack surface. Because SNNs rely on the precise accumulation of membrane potentials over time, attackers can manipulate the timing, sequence, and distribution of spikes to completely subvert model decisions without altering the spatial structure or overall intensity of the input signal.
Disrupting Information Integration via Spike Retiming
To understand why temporal attacks work, we must look at how an SNN integrates information. A spiking neuron accumulates incoming signals into an internal membrane potential over multiple time steps. If an attacker can alter when a spike arrives, they can disrupt this accumulation process, causing the neuron to fire incorrectly or fail to fire at all.
Crucially, this can be achieved without changing the total number of spikes. As demonstrated by Yi Yu et al. in Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks, an adversary can achieve high attack success rates by simply shifting the timing of existing spikes while keeping the total event count and intensity completely intact. This strategy rephases the input signals so that they conflict with the network's learned temporal filters. Because traditional defense mechanisms look for anomalies in signal intensity or spike rates, these rate-neutral rearrangements remain entirely stealthy.
Clean Stream: --|---|---|-----|--> (Membrane potential builds normally)
Retimed Stream: -|-----|---|---|---> (Altered phase misaligns with temporal filters)
This vulnerability to timing manipulation appears across various event-based processing frameworks:
- Alberto Marchisio et al. in DVS-Attacks: Adversarial Attacks on Dynamic Vision Sensors for Spiking Neural Networks show that precisely manipulating the sequence and timing of events directly disrupts the accumulation of membrane potentials, rendering standard noise filters insufficient for security.
- Hyun Myung et al. in Adversarial Attack for Asynchronous Event-Based Data establish that shifting existing events across the temporal boundaries of the network's underlying bins allows attackers to maintain data validity while successfully misleading the model.
- Even unintentional temporal variations can be destructive. Berken Utku Demirel et al. in Shifting the Paradigm: A Diffeomorphism Between Time Series Data Manifolds for Achieving Shift-Invariancy in Deep Learning note that small temporal delays function as inadvertent adversarial attacks, causing standard architectures to fail unless inputs are normalized into a canonical representation.
Temporal Trigger Splitting in Backdoor Attacks
The temporal dimension also fundamentally alters how backdoor attacks operate. In traditional vision models, a backdoor trigger is a spatial pattern stamped onto an image. In neuromorphic data streams, however, a trigger can be sliced and distributed across the temporal dimension.
This mechanism, explored by Gorka Abad et al. in Poster: Backdoor Attacks on Spiking NNs and Neuromorphic Datasets, relies on partitioning a malicious trigger across distinct, non-overlapping time frames. Because individual event-based frames have an incredibly low information density, these isolated fragments remain statistically inconspicuous and invisible to frame-by-frame inspection tools.
As the authors detail in Sneaky Spikes: Uncovering Stealthy Backdoor Attacks in Spiking Neural Networks with Neuromorphic Data and Time-Distributed Backdoor Attacks on Federated Spiking Learning, this temporal masking ensures that the backdoor only activates when the full sequence of spikes integrates over time at inference. In distributed learning environments like federated learning, multiple compromised devices can each contribute a tiny, clean-looking temporal fragment, allowing the global model to assemble a complete backdoor while individual updates evade detection.
The Vulnerability of Rate vs. Temporal Coding
Why are some SNNs highly vulnerable to these attacks while others show resilience? The answer lies in how the network encodes information across time.
Many high-performance SNNs rely on rate coding, where information is represented by the overall frequency of spikes rather than their exact timing. Jianhao Ding et al. in Rate Gradient Approximation Attack Threats Deep Spiking Neural Networks expose a critical flaw in this design: rate-coded networks behave similarly to continuous analog networks. This allows attackers to bypass complex temporal dynamics entirely and successfully compute adversarial perturbations using simple rate gradient approximations.
Conversely, the authors note that these rate-based approximations fail against networks utilizing precise temporal coding. To break these systems, hybrid approaches are required. Zecheng Hao et al. in Threaten Spiking Neural Networks through Combining Rate and Temporal Information demonstrate that attacks targeting both rate and timing information simultaneously achieve significantly higher success rates, proving that SNN vulnerability is often masked by a failure to exploit the network's underlying temporal dynamics.
The choice of network latency and resolution further dictates this vulnerability:
- High Latency vs. Low Latency: Massoud Pedram et al. in HIRE-SNN: Harnessing the Inherent Robustness of Energy-Efficient Deep Spiking Neural Networks by Training with Crafted Input Noise argue that the perceived robustness of SNNs is largely an artifact of high-latency rate-coding. When models shift to low-latency direct-coding for energy efficiency, their susceptibility to adversarial noise spikes dramatically.
- Temporal Resolution: Saima Sharmin et al. in Inherent Adversarial Robustness of Deep Spiking Neural Networks: Effects of Discrete Input Encoding and Non-Linear Activations identify an inverse relationship between temporal resolution and robustness. Increasing the number of timesteps helps clean accuracy but causes the network to mimic a continuous-valued model, expanding its vulnerability to gradient-based attacks.
Diagnosing and Defending the Temporal Dimension
To secure neuromorphic systems, researchers must analyze how gradients propagate across time steps. A primary mathematical driver of temporal vulnerability is Gradient Temporal Collinearity (GTC).
As defined by Desong Zhang et al. in Breaking Gradient Temporal Collinearity for Robust Spiking Neural Networks, GTC measures the redundancy of gradients across different time steps. When a network exhibits high gradient redundancy, it lacks temporal diversity, which correlates with a larger Hessian spectral radius and an increased sensitivity to adversarial noise. By breaking this collinearity, networks can be made significantly more robust.
High Collinearity (Vulnerable): Step 1 Gradient ≈ Step 2 Gradient ≈ Step 3 Gradient
Decorrelated (Robust): Step 1 Gradient ⊥ Step 2 Gradient ⊥ Step 3 Gradient
Defensive frameworks must directly account for these time-dependent dynamics rather than relying on static approximations. Jianhao Ding et al. in SNN-RAT: Robustness-enhanced Spiking Neural Network through Regularized Adversarial Training show that attacks utilizing backpropagation through time (BPTT) are highly effective at exploiting temporal dynamics, meaning defenses must incorporate regularized adversarial training that explicitly bounds these temporal gradients.
Furthermore, the unique properties of spike integration can be leveraged to build native defenses. The single-author work by Jianhao Ding (Neuromorphic computing paradigms enhance robustness through spiking neural networks) demonstrates that SNNs can outpace traditional ANNs in robustness if they prioritize task-critical information early in the spike sequence. By combining this prioritization with "early exit" decoding, the network can process the necessary information and stop before later-arriving adversarial perturbations have a chance to corrupt the membrane potentials.
Ultimately, understanding these temporal dependencies is vital not only for adversarial security but also for privacy. As Jiaxin Li et al. in Membership Privacy Evaluation in Deep Spiking Neural Networks point out, temporal statistics such as firing rates and average membrane potentials serve as direct proxies for confidence scores, meaning that the temporal dimension remains the primary vector for both attacking and auditing neuromorphic systems.
Key Foundational Work
DVS-Attacks: Adversarial Attacks on Dynamic Vision Sensors for Spiking Neural Networks (2021) — contributed 2 concept elements to this direction.
Inherent Adversarial Robustness of Deep Spiking Neural Networks: Effects of Discrete Input Encoding and Non-linear Activations (2020) — contributed 1 concept elements to this direction.
HIRE-SNN: Harnessing the Inherent Robustness of Energy-Efficient Deep Spiking Neural Networks by Training with Crafted Input Noise (2021) — contributed 1 concept elements to this direction.