Spiking neural networks (SNNs) process data the way event cameras produce it: as sparse streams of binary events, each carrying a pixel location, a polarity, and a timestamp. Because these networks accumulate evidence over many time steps, small changes in when events arrive can flip their predictions while leaving the total number of events untouched — an attack surface established by spike-retiming work that achieves over 90% attack success on DVS-Gesture while touching fewer than 2% of spikes Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks. In parallel, gradient inversion research has shown that in federated learning — where clients keep raw data local and send only parameter gradients to a server — those gradients leak the training inputs: an eavesdropper can reconstruct images by optimizing a dummy input until its gradients match the transmitted ones Gradient Inversion with Generative Image Prior.
Nobody has combined these two failure modes. When an SNN client trains by backpropagation through time and sends its gradients, the timestamps of its private event stream flow directly into the update signal — the very signal whose temporal sensitivity makes SNNs attackable. The one paper bridging neuromorphic data and federated learning, Time Bandits Time-Distributed Backdoor Attacks on Federated Spiking Learning, plants hidden triggers (an integrity attack, 100% attack success on N-MNIST) but never attempts to read private data back out of client updates. The closest privacy work on SNNs evaluates only membership inference — deciding whether a sample was in the training set, reaching 92.8% attack accuracy on CIFAR10-DVS for directly trained SNNs versus 84.3% for converted ANNs Membership Privacy Evaluation in Deep Spiking Neural Networks — which reveals a yes/no fact, not the content of the stream.
Core Claim
I propose to build and study SpikeInvert, the first gradient inversion attack that reconstructs binary spike trains (which pixels spiked, with which polarity, at which time step) from the surrogate gradients a federated SNN client transmits. The claim has two parts. First, inversion is feasible in principle: the same differentiable soft-relaxation machinery that makes retiming attacks powerful on the forward side of SNNs can be repurposed on the inverse side, optimizing a dummy event stream toward matching gradients. Second, and more surprising, temporal redundancy acts as an accidental privacy shield: when gradients across time steps are highly collinear (the condition the robustness literature calls Gradient Temporal Collinearity, GTC), individual time-step contributions superimpose and are hard to separate — so the very defenses that fix adversarial fragility may amplify privacy leakage. This tension between robustness and privacy has never been measured.
The mechanism in one sentence: because a spiking neuron accumulates inputs into a membrane potential over time, a gradient at the first layer encodes a weighted sum over time steps, and an attacker who can separate those overlapping contributions can recover when each event arrived — information that has provable adversarial value, since a single time-bin shift of existing spikes already drives attack success above 96% Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks.
Research Program
The attack formulates reconstruction as optimizing a dummy event tensor $S' \in \{0,1\}^{C \times H \times W \times T}$ (channels, height, width, time steps) so that its gradients match the intercepted ones:
$$\arg\min_{S'} \; \mathcal{D}\big(\nabla_W \mathcal{L}(S'), \; \nabla_W \mathcal{L}(S)\big) \; + \; \lambda_1 \|S'\|_1 \; + \; \lambda_2 \, \text{Cap}(S')$$
where $\mathcal{D}$ is an $L_1$ distance on gradients (magnitude matching matters for sparse signals, following the finding in TS-Inverse: A Gradient Inversion Attack Tailored for Federated Time Series Forecasting Models that $L_1$ matching captures signal magnitude better than $L_2$), $\|S'\|_1$ enforces event sparsity, and $\text{Cap}(S')$ enforces the capacity-1 constraint — no two spikes of the same pixel and polarity in the same time bin — a structural property of real event streams. Optimization proceeds with a projected-in-the-loop scheme borrowed directly from the retiming attack's toolkit Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks: a differentiable relaxation over spike-placement probabilities in the backward pass, a hard projection onto valid discrete schedules in the forward pass. That paper used this machinery to mislead an SNN; I invert the direction and use it to recover its input through the surrogate-gradient graph, which nobody has attempted.
The second workstream is the privacy–robustness tension. The STOD defense of Breaking Gradient Temporal Collinearity for Robust Spiking Neural Networks raises accuracy under PGD attack on CIFAR-10 from 0.03% to 36.72% by forcing temporal gradients into distinct orthogonal subspaces. But the entanglement analysis in Concealing Sensitive Samples against Gradient Leakage in Federated Learning shows that gradient reconstruction becomes easier exactly when individual contributions occupy low-entanglement, orthogonal subspaces. If both hold, decorrelation for robustness should simultaneously hand the inverter cleanly separated per-time-step signals. I will measure reconstruction fidelity as a function of measured GTC across vanilla, STOD-trained, and adversarially trained SNN clients — the last being the regime where clean accuracy is known to collapse from roughly 95% to 69–72% under retiming adversarial training Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks, making the full trade-off triangle (clean accuracy, robustness, privacy) explicit.
Finally, for large local batches where per-sample signals superimpose, I will combine two temporal aggregation axes: collecting gradients across federated rounds with robust aggregation, as TGIAs-RO does to reconstruct batches up to size 128 on ImageNet at 224×224 Temporal Gradient Inversion Attacks With Robust Optimization, and exploiting the intrinsic time axis of the SNN input itself. In TGIAs-RO the temporal dimension is only training rounds over static data; here both axes exist simultaneously, an unexplored regime flagged in the literature but never tested.
Baselines
The "do nothing" reference point is an eavesdropper running a plain DLG-style attack: optimize pixel-space dummy inputs with an $L_2$ gradient distance and a total-variation prior, ignoring the event structure entirely. This is what a naive attacker with existing tooling would do, and it establishes the floor; on binary sparse streams I expect it to fail because total-variation regularization actively degrades non-smooth signals, a result established for continuous time series in TS-Inverse: A Gradient Inversion Attack Tailored for Federated Time Series Forecasting Models. The strongest competitor is TS-Inverse itself, adapted to event streams, because it is the only existing inversion attack designed for temporal data rather than images: it brings quantile-bound prediction via an auxiliary model, trend and periodicity regularizers, and $L_1$ gradient matching. I will run it on the same SNN clients, same datasets, same intercepted gradients, with the same GPU-hours and the same hyperparameter search budget (grid sweeps on a held-out validation split, matched iteration counts for all attackers). Note honestly that no reported result found on Lacuna for TS-Inverse on any neuromorphic dataset; its reported 2x to 10x sMAPE improvement over DLG/InvG is on continuous forecasting benchmarks (Electricity 370, London Smartmeter, KDDCup 2018), so its neuromorphic performance is genuinely unknown — which is precisely why this comparison is decisive rather than rhetorical. A second strong competitor is GIAS, the generative-prior attack that reconstructs at batch size 32 with PSNR 15.58 dB on 64×64 ImageNet versus 13.78 dB for the pixel baseline Gradient Inversion with Generative Image Prior, testing whether a GAN prior trained on event-frame renderings substitutes for native event priors. Beating these baselines means strictly higher event-level timestamp F1 (spike location, polarity, and time bin all correct) at matched spike-count recall, across at least two of the three datasets, plus higher PSNR on rate-map renderings for continuity with the image-inversion literature. Two ablations isolate the proposal's key ingredients: removing the hard discrete projection (optimizing continuous relaxed spike intensities) isolates the projected-in-the-loop contribution, and removing multi-round aggregation isolates the TGIAs-RO-style temporal accumulation; a third, training the client SNN with STOD versus vanilla training, isolates the GTC–privacy claim.
Why not just…?
Why not just apply TS-Inverse or GIAS unchanged? This is the real threat, and I do not argue around it — it is the decisive experiment of the proposal. The evidence for a mismatch is indirect but consistent: total-variation priors assume piecewise-smooth signals and demonstrably smooth away the sharp volatility that defines time-series reconstruction TS-Inverse: A Gradient Inversion Attack Tailored for Federated Time Series Forecasting Models, and event streams are more extreme than forecasting signals — binary, extremely sparse, and capacity-constrained Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks. But if the adapted TS-Inverse recovers timestamps as well as SpikeInvert, the correct conclusion is that the attack side needs no new machinery, and I would report that and pivot the paper to the GTC–privacy tension, which no existing attack addresses either way.
Why not just add differential-privacy noise to the gradients? DP does block inversion, but the cost is documented: in Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks, DP-level security dropped MNIST task accuracy to 78–85% while deterministic mixed quantization kept 96.67% accuracy and collapsed DLG reconstruction PSNR from 55.3 to 8.1. Moreover, TGIAs-RO recovers private samples despite DP by aggregating gradients across rounds Temporal Gradient Inversion Attacks With Robust Optimization. SpikeInvert inherits this aggregation, so whether DP defends SNN clients at acceptable utility is an open empirical question I test directly rather than assume. The proposal's stance is that knowing how much leaks and at what defense cost is a prerequisite for choosing the defense.
Why not just require large local batches? Averaging over batch $B \ge 32$ is the cheapest mitigation, but the file's evidence rules it out as a complete answer: generative priors reconstruct at batch 32 with quality matching prior attacks at batch 4 Gradient Inversion with Generative Image Prior, and multi-round robust aggregation scales recovery to batch 128 Temporal Gradient Inversion Attacks With Robust Optimization. Batch scaling will still be one of my measured axes, and if event-level F1 collapses by, say, batch 8 on all datasets, that itself is a publishable negative result bounding the threat.
Why not just use membership inference to measure SNN privacy instead? Membership inference Membership Privacy Evaluation in Deep Spiking Neural Networks answers whether a sample was used, not what it contained. Its 92.8% versus 84.3% result on CIFAR10-DVS is evidence that directly trained SNNs memorize temporal structure, and it motivates this proposal; but a defender cannot reason about what an eavesdropper learns from gradients — timestamps, and ultimately the identity of the recorded subject, given that DVS-Gesture contains 29 identifiable human subjects A Low Power, Fully Event-Based Gesture Recognition System — from a binary membership signal alone.
Proposed Experiments and Decisive Checks
Datasets: N-MNIST, CIFAR10-DVS, and DVS-Gesture (1,342 gesture recordings, 11 classes, where the fully event-based reference system reaches 94.59% accuracy A Low Power, Fully Event-Based Gesture Recognition System). Clients: small CNN-style SNNs with leaky integrate-and-fire neurons trained by backpropagation through time with surrogate gradients, federated for 100 rounds. Metrics: event-level timestamp F1 (location, polarity, and time bin jointly correct), spike-count recall, PSNR on rate-map renderings for comparability with GIAS's 15.58 dB reference, and leaked-event attack success — feeding reconstructed timings into the retiming attack to show recovered timestamps are adversarially actionable, calibrating against its reported >90% attack success at under 2% tampered spikes on DVS-Gesture Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks.
Three checks decide the proposal. First, feasibility: SpikeInvert must beat the adapted TS-Inverse and GIAS baselines on timestamp F1 under matched compute; if it does not, the paper's contribution reduces to the negative finding plus the tension analysis. Second, the tension: reconstruction fidelity plotted against measured GTC across vanilla, STOD-trained, and adversarially trained clients. The hypothesis — that STOD's decorrelation raises leakage while restoring PGD accuracy from 0.03% toward 36.72% Breaking Gradient Temporal Collinearity for Robust Spiking Neural Networks] — must hold with a measurable fidelity gap; if fidelity is flat in GTC, the shield claim is false and I will report that. Third, the defense frontier: timestamp F1 under gradient-quantization and DP defenses at matched task accuracy, mapping where the robustness–privacy–utility triangle actually sits for federated neuromorphic systems. No reported result found on Lacuna for gradient inversion of any kind on neuromorphic data, so every number in the first two checks will be new.
References
- Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks
- Gradient Inversion with Generative Image Prior
- Time-Distributed Backdoor Attacks on Federated Spiking Learning
- Membership Privacy Evaluation in Deep Spiking Neural Networks
- TS-Inverse: A Gradient Inversion Attack Tailored for Federated Time Series Forecasting Models
- Breaking Gradient Temporal Collinearity for Robust Spiking Neural Networks
- Concealing Sensitive Samples against Gradient Leakage in Federated Learning
- Temporal Gradient Inversion Attacks With Robust Optimization
- Mixed Quantization Enabled Federated Learning to Tackle Gradient Inversion Attacks
- A Low Power, Fully Event-Based Gesture Recognition System