Many evolving networks, such as email logs, purchase histories and contact traces, are sensitive, and models that learn from them are increasingly expected to run continuously on low-power devices. Two literatures meet here without speaking to each other. Differentially private graph learning protects individual edges by adding calibrated noise, but every new snapshot of a changing graph spends more of a fixed privacy budget, so utility collapses as the stream grows. Spiking neural networks (networks whose neurons accumulate input and emit a binary pulse only when an internal "membrane potential" crosses a threshold) save energy because computation happens only when a neuron fires. Our proposal is that these two problems share one solution. A neuron that stays silent until its input has changed enough is also a privacy mechanism that pays budget only when it speaks.
Concretely, we build a dynamic graph network whose only graph-touching layer is made of leaky integrate-and-fire neurons with randomized thresholds. These neurons are designed so that each emitted spike is an output of the Sparse Vector Technique, the classical differential-privacy method that answers "is this above threshold?" and charges privacy cost only for "yes" answers. Everything downstream, including spiking temporal layers, learnable delays and classifiers, only sees those spikes, and so it is free post-processing. The number of spikes then sets both the privacy loss and the energy bill, and a single threshold controls both. The goal is a model whose privacy cost grows with how much the graph actually changes, rather than with the number of timesteps.
Core Claim
Under event-level edge differential privacy (two graph streams count as neighbors if they differ by one edge insertion or deletion event), a spiking first layer with noisy, re-randomized thresholds and a hard per-node spike cap over a sliding window of $w$ steps has privacy loss bounded by the number of spikes the two endpoint nodes may emit in that window. The bound does not depend on the stream length $T$. On dynamic node classification benchmarks with long horizons (Tmall has 186 steps), this should give better accuracy at a fixed total privacy budget than releasing a noisy representation at every snapshot. It should also keep spike sparsity of the kind dynamic spiking graph networks already report, for example the 83–88% sparsity of ChronoSpike. What is new is the identification of firing with paying for privacy, the accounting that makes this identification sound, and an empirical test of whether the spiking form adds anything beyond a plain threshold-gated release.
Mechanism
Following the decoupling idea of GAP, we encode each node's own features with an encoder that never sees the graph. GAP perturbs multi-hop aggregates of one static graph, caches them, and trains on the cache. We reuse its separation of graph access from learning, but GAP's cache is invalidated at every edge update, which is exactly the budget problem we target. At time $t$ we form node $v$'s neighborhood aggregate $a_v(t)$, with each neighbor's contribution clipped to norm $C$. Each of $d$ gate neurons for node $v$ integrates the change in its channel:
$$U_v(t) = \lambda U_v(t-1) + \big(a_v(t) - a_v(t-1)\big) + \eta_v(t), \qquad s_v(t) = \mathbf{1}\big[U_v(t) \ge V_{\text{th}} + \rho_v\big],$$
Here $\lambda$ is the membrane leak, $\eta_v(t)$ is fresh Laplace noise added to each query, $\rho_v$ is threshold noise that is resampled after every spike, and the potential is soft-reset by $V_{\text{th}}$ on firing. One edge event shifts any node's membrane trajectory by at most a constant multiple of $C$, because the leaky integral of differences of clipped aggregates is bounded. That is the bounded-sensitivity query stream the Sparse Vector Technique requires. Each neuron then behaves as an instance of that technique, and capping spikes at $c$ per node per window yields a window-level bound of roughly
$$\varepsilon_w \lesssim 2\,c\,\varepsilon_s,$$
where $\varepsilon_s$ is the per-spike cost set by the two noise scales, and the factor 2 counts the edge's two endpoints. This is a target to be proved carefully, not a result we assume. Lyu et al. showed that many published variants of the technique leak without bound because of subtle accounting errors, which is why we fix the variant (noise on every query, fresh threshold noise after every positive answer) and audit it empirically. The window formulation mirrors the $w$-event guarantee of WGT-LDP, which caps how often each node is sampled for continuous graph publication. We get a comparable cap from the neuron's own refractory budget instead of a sampling scheduler.
Downstream of the gate we place a dynamic spiking graph backbone. We use Delay-DSGN, which adds learnable Gaussian synaptic delays to leaky integrate-and-fire layers, and Dy-SIGN, which trains with constant memory by implicit differentiation at an equilibrium state. Neither has any privacy mechanism. Because these layers only consume private spikes, their learned delays cannot leak more than the spikes do. This answers by construction the worry that per-edge delays act as a temporal side channel, the kind of temporal signature that Mauw et al. and Sun et al. show breaks snapshot anonymization. Multi-hop aggregation over the graph needs further gated layers, and each extra hop composes privacy cost. We expect one or two hops to be the practical limit, so we compensate for shallow aggregation using the encoder and temporal depth rather than the 64-hop averaging that UPGNET uses for static locally private features.
Research Program
The work has three parts. The first is theoretical. We prove the window-level bound for leaky, soft-reset neurons with a spike cap, including the multi-channel composition and the K-hop extension, and state it with Rényi or zCDP accounting for tighter totals. The second is the system: the gate layer plus a Delay-DSGN or Dy-SIGN backbone, trained on the spike outputs with ordinary (non-private) optimization. Labels and node features are treated as public, as in GAP's edge-level setting; node-level privacy is out of scope. We also check whether Dy-SIGN's fixed-point convergence survives noisy spike inputs, since noise could break the contraction conditions its gradients rely on. The third is measurement. Energy is reported as spike counts and as accumulate and multiply-accumulate operation counts per inference, converted with standard per-operation energy estimates. Deployment on neuromorphic hardware, where such counts translate into the two-to-three-orders-of-magnitude savings that Martí et al. report for TrueNorth, is a stretch goal rather than a dependency.
Baselines
The naive reference point is a non-private Delay-DSGN or Dy-SIGN on raw streams. It sets the utility ceiling and, with edge-inference attacks, the leakage floor. GAP reports that non-private models allow membership attacks up to 81% AUC versus 50–52% under GAP, and we expect a comparable gap.
The strongest existing private learning method is GAP applied per snapshot, with the total budget split as $\varepsilon/T$ under sequential composition, feeding the same spiking backbone. It is the strongest competitor because it is the best-documented edge-level DP graph learner: 76.3% accuracy on Facebook at $\varepsilon = 4$ under edge-level DP, versus 50.4% for the next private GNN, as reported in GAP. We also run a stronger variant that spends budget only on snapshots whose degree change exceeds a public schedule.
The strongest data-side competitor is WGT-LDP. It publishes private synthetic weighted snapshots, and Delay-DSGN is then trained on them as post-processing. This is the most dangerous baseline because it already handles continuous release.
A weight-level baseline applies DP-SGD to the dynamic spiking network, as DPSNN does for images. It protects the trained weights but not the spike stream at inference, so we report it with its narrower guarantee stated explicitly.
All methods share the same encoder, backbone, hyperparameter search budget (an equal number of trials, tuned on non-private validation proxies to avoid charging tuning to privacy), GPU hours, and privacy accountant. Beating the baselines means higher Macro-F1 than per-snapshot GAP and WGT-LDP at equal $\varepsilon_w$ on at least Tmall, where $T$ is largest, without a higher operation count. No absolute Macro-F1 values for DBLP, Tmall or Patent, private or not, were found on Lacuna. The file gives only relative claims, such as ChronoSpike's 2.0% Macro-F1 gain over twelve baselines and Delay-DSGN's more-than-2% gain over Dy-SIGN on DBLP, so we will reproduce all baselines ourselves.
Two ablations isolate the key ingredient. The first, a gate without threshold, adds Gaussian noise to the aggregate change at every step and calibrates it to the same total budget. It tests whether threshold gating is what saves budget. The second, a gate without spikes, keeps the identical Sparse Vector gate but releases a continuous noisy message when it fires, into a continuous dynamic backbone. It tests whether binarization costs utility, or saves energy, relative to the same privacy logic.
Why not just…?
Why not just run GAP or UPGNET on a low-power quantized GPU? Hardware does not fix the accounting. Any per-snapshot private release composes over $T$ steps, and Zhang et al. and Korolova et al. document that this drives continuous private analytics toward useless noise, sometimes to the level of random guessing. Quantization reduces energy per operation but still runs every operation at every step. The per-snapshot GAP baseline measures the accounting half of this argument directly. The energy half is only estimated from operation counts unless we reach hardware.
Why not just apply a Sparse Vector gate to an ordinary continuous dynamic GNN? This is the real threat. The privacy saving comes from the threshold, not from biology, and it is quite plausible that a continuous gated model matches our accuracy. We have therefore framed the claim around the gate. Spiking is the implementation in which the same gate also removes multiplications and makes idle nodes free. The no-spike ablation settles the question. If it matches us on F1 at equal $\varepsilon_w$, the contribution narrows honestly to "the private gate costs nothing extra in spiking form and buys sparsity". If it beats us, binarization noise is hurting, and we will report that.
Why not just privatize the data once with WGT-LDP and train anything on the output? Post-processing is free, so this is legitimate and strong. Its weakness, noted in the WGT-LDP paper itself, is that non-sampled nodes inherit stale states, which drifts when the graph changes quickly. Our gate fires exactly when change accumulates. Comparing the two on Tmall (fast-changing) versus Patent (slow), and on WGT-LDP's own Synthetic-I and Synthetic-II streams (20% versus 90% snapshot change), should show where each wins. A WGT-LDP win on slow graphs would be an acceptable and informative outcome. A caveat is that WGT-LDP gives local DP, which is a stronger trust model than our central edge-level DP, and we will report the comparison with that caveat.
Why not just use homomorphic encryption with spiking networks? FHE-DiSNN takes 8–16 seconds per MNIST inference on a shallow network, which is far from streaming graphs. Encryption also protects computation, not what the outputs reveal about edges, so it would not replace DP even if it were fast enough.
Proposed Experiments
We evaluate dynamic node classification on DBLP (27 steps), Tmall (186 steps) and Patent, reporting Macro-F1 and Micro-F1 over $\varepsilon_w \in \{0.5, 1, 2, 4, 8\}$ and window sizes $w$. As a static calibration check, we run on Facebook with $T = 1$, where our method reduces to a thresholded GAP and should land near GAP's reported regime at $\varepsilon = 4$.
The first decisive check plots F1 against stream length at fixed total budget, by truncating Tmall to increasing $T$. Per-snapshot GAP should degrade steadily as $T$ grows, while our curve should be governed by the graph's change rate. If the two curves degrade alike, the core claim fails.
The second check sweeps the threshold and plots F1, $\varepsilon_w$ and operations per inference together, to test whether a single knob really traces a joint privacy–energy–utility frontier.
The third is a privacy audit. We insert canary edges, run edge-inference and membership attacks on the spike outputs and on learned delays, and estimate empirical $\varepsilon$ with a distinguishing test. Any audited value above the claimed bound means the accounting is wrong.
The fourth tests stability. We measure fixed-point convergence rates and gradient norms of the Dy-SIGN backbone as the gate's noise scale increases, to see whether noisy spike inputs destabilize equilibrium training.
Together with the two ablations, these experiments determine whether spiking adds anything beyond the threshold gate. They also determine whether the threshold gate beats both per-step noise and one-shot private publication.
References
- ChronoSpike: An Adaptive Spiking Graph Neural Network for Dynamic Graphs
- GAP: Differentially Private Graph Neural Networks with Aggregation Perturbation
- Understanding the Sparse Vector Technique for Differential Privacy
- Continuous Publication of Weighted Graphs with Local Differential Privacy
- Delay-DSGN: A Dynamic Spiking Graph Neural Network with Delay Mechanisms for Evolving Graph
- Dynamic Spiking Graph Neural Networks
- Preventing active re-identification attacks on social graphs via sybil subgraph obfuscation
- Privacy Preserving Social Network Publication Against Mutual Friend Attacks
- Going Deeper into Locally Differentially Private Graph Neural Networks
- Energy-Efficient Neuromorphic Classifiers
- DPSNN: A Differentially Private Spiking Neural Network with Temporal Enhanced Pooling
- Structure-Preference Enabled Graph Embedding Generation Under Differential Privacy
- Personalized Social Recommendations - Accurate or Private?
- Privacy-Preserving Discretized Spiking Neural Networks