Event cameras report only pixels whose brightness changed, and spiking neural networks (SNNs) process those events as sparse binary pulses. Each pulse costs energy, so fewer pulses means less energy. This pairing is often sold as private by construction, because a person standing still vanishes from the stream and fine facial texture is never recorded. The evidence says otherwise. Timing survives, and timing identifies people: the Bullying10K authors concede that gait remains embedded in event dynamics, and SNNs trained on event data leak training-set membership more than conventional networks do. We propose a privacy mechanism that sits inside the sensor, before any spike reaches the network or leaves the device. It randomly drops events and randomly re-times the survivors within a bounded window. The same drop probability that sets the formal privacy guarantee also sets how many spikes the downstream SNN must process. Privacy and energy therefore improve together rather than trading off.
The claim is testable and concrete. On event-camera action recognition, a sensor-side mechanism that thins and jitters events at random should push a gait-based re-identification attacker toward chance at a lower task-accuracy cost than doing nothing, coarsening time, or training the network against an identity adversary. It should also cut synaptic operations roughly in proportion to the keep rate, and it should come with a per-event guarantee that none of those alternatives can state.
Core Claim
Existing privacy work on SNNs protects the model, not the timing of the input. DPSNN adds differentially private gradient noise during SNN training, using clipped gradients plus Gaussian noise. That protects membership of training examples but says nothing about what a deployed stream reveals about the person in front of the camera. PrivateSNN uses temporal fine-tuning to scramble converted weights against class reconstruction on static images, with no formal guarantee. SpyKing encrypts SNN inference homomorphically, so computation runs on encrypted data, at 15 minutes per image. On the trajectory side, Differentially Private Timestamps Publishing in Trajectory and Efficient Frequency-Based Randomization privatize timestamps and visit frequencies in offline GPS databases, and both note that they do not handle streams.
Our proposal differs from all of these in where the noise goes, which is the event stream at inference time. It also differs in what the noise costs. Every mechanism above adds computation or noise on top of the pipeline. Ours removes spikes, so the privacy budget is paid in spikes that were going to cost energy anyway.
Mechanism and Threat Model
The deployment we have in mind is an event camera plus SNN monitoring a public space, as in Bullying10K. The adversary observes the stream the device emits, either raw events or first-layer spikes forwarded to a server. The adversary wants to re-identify individuals from their movement rhythm. The honest task is action recognition.
Our mechanism works on each event in three steps. The event first has its timestamp coarsened to a bin. It is then kept with probability $q$. If kept, it is assigned an output bin drawn from a two-sided geometric distribution centered on its true bin and truncated to a window of $\pm W$ bins. Truncation keeps displacement bounded and avoids the temporal inversions that unbounded Laplace noise causes, a failure mode the trajectory literature documents. We will also enforce the one-spike-per-pixel-per-bin rule used in Time Is All It Takes.
Define two streams as neighbors when a single event differs in presence or in timing within the window. The re-timing step alone gives per-event local differential privacy $\varepsilon_0$, set by the geometric decay rate, where local means each event is privatized before anyone else sees it. Random dropping then amplifies the guarantee, by the standard subsampling argument for local mechanisms, to
$$\varepsilon(q) \approx \log\left(1 + q\,(e^{\varepsilon_0}-1)\right),$$
where $\varepsilon(q)$ is the effective per-event guarantee. For the energy side, the expected input spike count is multiplied by $q$, so first-layer accumulate operations scale by $q$. Accumulate operations are the additions an SNN performs in place of multiplications, and Tavanaei et al. and Jang et al. tie dynamic SNN energy to spike count.
We state the limitation of the guarantee plainly. It covers individual events, and gait is a pattern spread over thousands of events. Composition over a whole clip gives a weak formal bound. The formal guarantee is therefore a floor on fine-timing protection, and the claim about gait must be settled empirically with strong attackers. That is where most of the experimental effort goes.
Two findings motivate why the network should tolerate this noise. DPSNN found that the decay in leaky integrate-and-fire neurons, whose internal state slowly forgets old input, discards injected noise better than neurons without decay. SpyKing found that integration and thresholding in these neurons filter encryption noise that collapses conventional networks.
One finding constrains the design. Spike-retiming attacks show SNN outputs can be hypersensitive to timing, with over 90% attack success when fewer than 2% of spikes are moved. That result concerns adversarially chosen shifts, not random ones. It still warns that an SNN trained on clean streams may break under our mechanism, so we always train with the mechanism in the loop.
Research Program
The work has three parts. First, we build the mechanism as a streaming, per-event operator that can live at the sensor front end, and we train SNNs with it active. Candidate backbones are spiking convolutional networks in SpikingJelly and the spiking transformer QKFormer.
Second, we build attackers strong enough to keep the evaluation honest. These are a gait re-identification network trained on clean streams, a second one retrained on privatized streams (an adaptive attacker), and a denoising attacker that first tries to undo the jitter. We also run the membership-inference protocol of Membership Privacy Evaluation in Deep SNNs, which substitutes firing rates and membrane potentials for output probabilities, as a secondary leakage measure.
Third, we trace privacy–utility–energy curves by sweeping $q$, $W$ and $\varepsilon_0$. Energy is measured as counted synaptic operations. Where hardware access permits, we also measure it on Loihi, the platform on which Tang reported over two orders of magnitude gains over embedded GPUs.
Baselines
The naive reference point is an unprotected SNN on the raw stream. It fixes the upper end of accuracy, energy and leakage.
The strongest cheap competitor is deterministic temporal coarsening: fewer, wider time bins, or deterministic decimation of events. It reduces energy and, per Li et al., fewer time steps reduce membership leakage. We run it at matched spike counts and matched accuracy.
The strongest empirical-defense competitor is adversarial privacy training, in which the encoder is penalized when a simultaneously trained identity classifier succeeds. It gets the same backbone, data and tuning budget as our method, and its identity head is as strong as our attacker. The spike-retiming paper reports that adversarial training on DVS-Gesture drops clean accuracy from about 95% to about 69–72%. That figure is for robustness to timing attacks, not privacy, so it serves only as a warning about cost, not a target.
We also run EventDrop augmentation, which Li et al. found only partially mitigates membership leakage, and DPSNN-style private training, which answers a different question but is the established differential-privacy method for SNNs. For DPSNN, the file reports 43.24% on CIFAR10-DVS under DP. As a timestamp-noise reference we port truncated Laplace timestamp perturbation from Yan et al.
Two ablations isolate the key ingredient. Thinning-only and jitter-only separate the two noise sources. Deterministic decimation at identical spike count tests whether randomness, rather than simply having fewer spikes, is what defeats the attacker.
We count as winning an equal or lower adaptive-attacker re-identification rate than every baseline at equal or lower synaptic operations, with task accuracy within a few points of the unprotected SNN. We found no reported result on Lacuna for gait re-identification from event streams, so the unprotected attacker accuracy will itself be a new measurement.
Why not just…?
Why not just coarsen time or use fewer time steps? This is the most serious threat, because it is free and also saves energy. If deterministic coarsening at matched spike count pushes re-identification as low as our mechanism at similar accuracy, the random component adds little beyond a paper guarantee. We would then report that honestly and reposition the contribution as certification rather than empirical gain. Our expectation that it falls short rests on how gait is carried: in stride periodicity and relative phase between limbs. Coarsening preserves these at bin resolution and preserves them identically on every pass, so an attacker can learn them. The matched-count ablation settles this directly.
Why not just train the network adversarially against an identity classifier? It can look strong against the attacker it was trained with and gives no guarantee against a new one. It also protects only the network's outputs, not the raw stream that leaves the sensor. The retrained, adaptive attacker in our evaluation tests whether its protection survives. Its accuracy cost is measured under the same tuning budget as ours.
Why not just rely on EventDrop or DP-SGD during training? Both act on training. EventDrop leaves inference streams untouched, and Li et al. report that it does not remove membership leakage. DPSNN bounds what the weights reveal about training examples, not what a live stream reveals about the person being filmed. We include both so the difference in threat model shows up in the numbers.
Why not just encrypt? SpyKing's 15-minute single-image inference, and its need to decrypt intermediate states, rule out real-time edge use.
Proposed Experiments and Decisive Checks
The primary testbed is DVS-Gesture. We would use its per-subject recordings for identity labels, assuming the subject metadata in the original release is usable; this still needs to be confirmed. Its gesture labels serve as the task, and the spike-retiming paper provides reference accuracy of about 95%.
Bullying10K is the realistic surveillance setting, with its reported DVS action accuracy of about 70% and X3D's 76.90%. Identity experiments there depend on whether performer metadata exists, which we must verify before committing. CIFAR10-DVS is used only for the membership-inference check against the 92.8% SNN attack accuracy reported by Li et al.
As a transfer check, we apply the timestamp mechanism to Geolife check-in times and compare it with TDP, which reports an MSE of 140, a 6.7% improvement over its prior baseline.
The decisive result is a Pareto comparison. The randomized mechanism should beat deterministic decimation at matched spike count on adaptive re-identification while holding task accuracy. Membership-inference accuracy should fall as $q$ decreases. First-layer synaptic operations should track $q$ within measurement error.
If random thinning buys no privacy beyond deterministic thinning, the central claim fails.
References
- DPSNN: A Differentially Private Spiking Neural Network with Temporal Enhanced Pooling
- PrivateSNN: Privacy-Preserving Spiking Neural Networks
- SpyKing—Privacy-preserving framework for Spiking Neural Networks
- Differentially Private Timestamps Publishing in Trajectory
- Efficient Frequency-Based Randomization for Spatial Trajectories Under Differential Privacy
- Bullying10K: A Large-Scale Neuromorphic Dataset towards Privacy-Preserving Bullying Recognition
- Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks
- Deep learning in spiking neural networks
- Spiking Neural Networks - Part I: Detecting Spatial Patterns
- Membership Privacy Evaluation in Deep Spiking Neural Networks
- Deep Reinforcement Learning with Population-Coded Spiking Neural Network for Continuous Control