Spiking Neural Networks (SNNs) are often celebrated as the "third generation" of neural networks, offering a path toward energy-efficient, brain-like "Edge Intelligence" Shuiguang Deng (Edge Intelligence with Spiking Neural Networks). Unlike traditional Artificial Neural Networks (ANNs) that process static frames of continuous values, SNNs communicate via discrete, binary "spikes" over time.
However, this temporal dimension—the very thing that makes SNNs efficient—introduces a unique and stealthy security vulnerability. While traditional adversarial attacks focus on changing pixel values (spatial perturbations), attacks on SNNs can succeed by simply manipulating when a spike occurs.
The Temporal Attack Surface
In a standard image classifier, an adversary might change a pixel from 0.8 to 0.1 to fool the model. In an SNN, information is often encoded in the rate (how many spikes occur) or the latency (the precise timing of those spikes).
As Alberto Marchisio and colleagues demonstrate in DVS-Attacks, neuromorphic systems like Dynamic Vision Sensors (DVS) operate in a multidimensional spatio-temporal search space. Because SNNs integrate these sparse events over time, an attacker does not need to overwhelm the system with noise. Instead, they can inject a few carefully timed pulses that mimic legitimate sensor noise, bypassing standard spatial filters while still influencing the network's internal state.
Why Timing Matters: Membrane Potential
To understand why these attacks work, we must look at the fundamental mechanism of a spiking neuron: the membrane potential. A neuron accumulates incoming signals over time; once its internal potential crosses a specific threshold, it fires a spike and resets.
This accumulation process creates a "temporal buffer." Roberto Riaño and Gorka Abad found that SNNs are significantly more sensitive to triggers placed at the beginning or middle of an input sequence than at the end (Flashy Backdoor). Early spikes have more time to influence the accumulation of membrane potential, effectively "steering" the neuron toward a specific firing state before the legitimate signal can be fully processed.
Spike-Retiming: The Stealthy Jitter
The most counterintuitive finding in this research area is that an adversary doesn't even need to add or remove spikes to cause a misclassification. Yi Yu and Kun Wang introduced Spike-Retiming Attacks, which achieve over 90% success rates by shifting the timing of existing spikes by tiny increments (Time Is All It Takes).
Because the network's filters are tuned to precise phase alignments, shifting fewer than 2% of the spikes—without changing the total spike count—can trick the model. This makes the attack invisible to rate-based detection systems, which only look at the "volume" of activity rather than its "rhythm."
Distributed Backdoors in Time
The temporal nature of SNNs also enables a new class of "smart" backdoors. In traditional vision models, a backdoor trigger is usually a static patch (like a small square in the corner of an image). In neuromorphic data, Gorka Abad and Stjepan Picek show that triggers can be distributed across the temporal dimension (Sneaky Spikes).
This "Time Bandit" approach splits a malicious trigger into discrete time slices. No single frame contains enough information to be flagged as suspicious. However, because the SNN naturally integrates information over time, it reconstructs the full malicious trigger internally. This is particularly dangerous in Federated Learning environments, where Hanqing Fu notes that different malicious clients can contribute non-overlapping temporal segments of a trigger, making the attack nearly impossible to detect at the local level (Spikewhisper).
Moving Toward Temporal Robustness
If SNNs are uniquely vulnerable to time-based attacks, how can we defend them? Current research suggests that standard defenses are insufficient because they ignore the temporal evolution of a neuron's state Jianhao Ding (Robust Stable Spiking Neural Networks).
Promising directions include:
- Hybrid Attacks and Defenses: Zecheng Hao and Zhaofei Yu argue that defenses must simultaneously account for rate and temporal information (HART).
- Homeostasis: Hejia Geng and Peng Li propose using adaptive firing thresholds that dynamically adjust to abnormal activity, mimicking biological homeostasis to maintain stability under perturbation (HoSNN).
- Temporal Gradients: Training networks using gradients derived directly from spike timing, rather than converting pre-trained static models, can produce weight distributions that are inherently more resistant to temporal manipulation (Spike timing reshapes robustness against attacks in spiking neural networks).
By shifting the focus from spatial patterns to temporal dynamics, researchers are beginning to map the true security boundaries of the neuromorphic landscape.