Encryption hides what you send, but not how you send it: packet sizes, directions, and inter-arrival times still reveal which website you visit or which app you run, and deep classifiers exploit this with over 90% accuracy on a single benchmark capture. The known, documented failure is that these classifiers collapse when the capture conditions change — a different client CPU, a different network path, or a collection five years later drops accuracy from over 90% to under 60% (Seamless Website Fingerprinting in Multiple Environments). Meanwhile, a striking result from the spiking-neuron literature suggests a cheap fix: a 100-unit spiking network classifies encrypted traffic with 95.9% accuracy using only 31.5k parameters, ten times fewer than the CNN it beats (Encrypted internet traffic classification using a supervised spiking neural network). A spiking neuron — a unit that accumulates input over time and fires a discrete pulse when its internal state crosses a threshold, then resets — has one parameter, the leak, that is literally a temporal-scale knob. And a separate line of work showed that when the temporal resolution of the input changes, the leak can be recalibrated by a closed-form transformation, restoring a collapsed 38.8% transfer to 93.6% with no retraining at all (Zero-Shot Temporal Resolution Domain Adaptation for Spiking Neural Networks).
My claim is that these three threads belong together. Most cross-environment shift in traffic captures is, at first order, a change in the effective temporal scale of packet dynamics — jitter stretches inter-arrival times, slow CPUs space out packet generation, throttled circuits merge bursts. If that is right, then a small spiking traffic classifier should transfer across datasets by recalibrating its leak — estimated from a handful of unlabeled traces from the target environment — rather than by adversarial retraining over many labeled environments. The proposal builds exactly this system, measures where the linear-recalibration story holds and where it visibly breaks, and treats cross-dataset validation as the evaluation lens rather than an afterthought.
Core Claim
A large fraction of the cross-environment collapse in traffic classifiers can be absorbed by re-parameterizing the temporal dynamics of a small spiking network, without target labels and without retraining. Concretely, a decoupled spiking neuron (DA-LIF, which separates spatial input accumulation from temporal decay) updates its state as
$$u_t = \beta\, u_{t-1} + \alpha\, x_t, \qquad \text{fire if } u_t \geq \theta \text{, then reset } u_t \leftarrow 0,$$
where $x_t$ is the binned packet stream, $\alpha$ scales burst-volume accumulation, and $\beta$ is the leak. The resolution-ratio transformations of Zero-Shot Temporal Resolution Domain Adaptation for Spiking Neural Networks tell us how $\beta$ should change when the input's effective time step is rescaled by a factor $\rho$; the open problem is that in traffic, unlike speech, $\rho$ is unknown, non-uniform, and contaminated by non-linear effects (threshold resets, bufferbloat, dummy-packet injection). The research program is to estimate $\rho$ from unlabeled target traces by matching burst-duration and inter-arrival distributions between source and capture, validate which of the three published transformation formulations (Integral, Euler, Expectation) fits traffic, and honestly bound where the linear story fails.
This is deliberately different from the closest existing work. The supervised SNN traffic classifier (Encrypted internet traffic classification using a supervised spiking neural network) never leaves in-distribution splits and uses static leak rates; it supplies the architecture and the replication targets, not the question. The zero-shot SSM paper supplies the mathematics but was evaluated only on speech and event-camera data, and its authors flag non-linear resets as its known breakdown. Neither has ever faced network jitter, hardware side channels, or active padding defenses.
Research Program
Three components. First, build the classifier: a one-to-three-layer decoupled-LIF network over packet streams (direction, size, inter-arrival bins), trained with surrogate-gradient backpropagation through time exactly as in the existing spiking traffic work, targeting the same 30k-parameter regime. Second, build the recalibration: an estimator that, given unlabeled target traces, infers the effective temporal rescaling per capture environment by aligning source and target burst-duration statistics, then transforms $\beta$ (and, where the decoupled parameterization helps, re-weights $\alpha$) using the SSM-derived rules. Third, run a pre-registered failure analysis: conditions under which recalibration should degrade — rising firing rates (where the linear approximation degrades), and active defenses like WTF-PAD, which insert packets non-linearly and should not be expressible as any rescaling of $\beta$.
Baselines
The do-nothing reference is the same spiking classifier trained on a source environment and applied to the target with no adaptation; the analogous collapse for a CNN is documented at over 90% in-distribution falling to under 60% cross-environment (Seamless Website Fingerprinting in Multiple Environments), and whether a tiny LIF network is more or less fragile is precisely the open question this proposal answers. The strongest existing competitor depends on the benchmark: on the 8-environment transfer it is WFNet with a Gradient Reversal Layer, which recovered cross-environment accuracy to nearly 80% but required adversarial training with labeled data from many environments; on the five-year temporal drift it is NetCLR, which reached 72.1% on the Drift dataset versus 51.0% for Triplet Fingerprinting (Realistic Website Fingerprinting By Augmenting Network Traces); under active defense it is Deep Fingerprinting, which retains 90.7% against WTF-PAD where classical feature methods fall below 70% (Deep Fingerprinting). All baselines run on identical splits with an equal tuning budget, and the bar is explicit: the spiking classifier with zero-shot recalibration must come within two points of the competitor's published cross-domain number, or exceed it, at no more than a tenth of the parameters and with no target labels — otherwise the honest outcome is that the mechanism lies elsewhere. The key ablation isolates the ingredient twice: a static-leak coupled-LIF version (as in the existing traffic SNN) separates the value of decoupled-and-recalibrated dynamics, and an oracle-$\rho$ condition separates the recalibration math from the $\rho$ estimator. One further ablation I consider essential: apply the estimated temporal rescaling to the input stream of the CNN baseline instead of the leak of the spiking net. If the rescaled-input CNN matches the recalibrated SNN, the interesting mechanism is input rescaling, not spiking — and I would report it that way.
Why not just…?
Why not just use a CNN with a Gradient Reversal Layer? This is the real threat, and it already sits near 80% cross-environment. But it needs labeled training data from many environments gathered before deployment, plus adversarial training, whereas the proposal needs only unlabeled traces from the one target capture at deployment time. The head-to-head experiment on the 8-environment dataset, plus the input-rescaling ablation above, settles whether the spiking formulation contributes anything beyond what any classifier gains from a temporally rescaled input.
Why not just augment the training data? NetAugment lifted 5-shot transfer to 80% and holds 72.1% across five years of drift (Realistic Website Fingerprinting By Augmenting Network Traces), and Harmonious Data Augmentation lifted 5-shot Var-CNN from 17.9% to 59.7% on CW100 (Few-Shot Website Fingerprinting Attack). But augmentation must anticipate the perturbation family at training time; hardware and sampling-rate shifts at deployment are not in NetAugment's Tor-burst vocabulary, whereas parameter recalibration is zero-shot and compositional. Since the two are complementary rather than exclusive, the experiments include the combination (spiking net plus NetAugment-style augmentation) rather than arguing the alternative away.
Why not just use a pre-trained transformer? ET-BERT reaches 98.9% F1 on ISCX-VPN and 91.55% with 10% labels, but requires pre-training on 30GB of unlabeled traffic (ET-BERT); that is a different deployment regime from a line-rate edge tap, which is the regime where a 31.5k-parameter event-driven network matters, and where neuromorphic quantization results show real hardware savings (A New Learning Method for Inference Accuracy, Core Occupation, and Performance Co-optimization on TrueNorth Chip). If accuracy at any cost were the only question, ET-BERT wins; it is not the question here.
Why not classical features with a random forest? k-fingerprinting reaches an 85% true positive rate at 0.02% false positives on a 100,000-page open world (k-fingerprinting), but hand-crafted feature methods fall below 70% against WTF-PAD and depend on page-boundary detection, which the seamless-window setting removes. It stays in the comparison as the open-world scalability reference, not as the accuracy competitor.
Proposed Experiments
Sanity replication first: reproduce the spiking baseline at its published numbers on ISCX (95.9% on the 14-class task, 98.6% on Tor) before any transfer work. The main experiments then follow the shift taxonomy. Environment shift uses the 8-environment, 40-site dataset of Seamless Website Fingerprinting in Multiple Environments: train on one environment, transfer to the other seven, report top-1 accuracy against the under-60% naive and nearly-80% GRL references. Temporal drift uses AWF (2017) to Drift (five years later), against NetCLR's 72.1%. Defense shift uses the DF95 setting with and without WTF-PAD, with the pre-registered expectation that recalibration helps less here and the non-linear breakdown becomes the finding rather than the embarrassment. Protocol shift tests the shared functional classes (VoIP, chat, video, file transfer, browsing) between ISCX VPN-nonVPN and ISCX Tor-nonTor. Metrics throughout: top-1 accuracy, open-world TPR at fixed FPR, parameter count, and estimated spike-operation energy for the edge-deployment argument. The decisive check is a single plot: recovered accuracy versus estimated $\rho$ across all transfers, with the oracle-$\rho$ curve as the ceiling — if the estimated-$\rho$ curve tracks the oracle, the core claim stands; if it does not, the estimator, not the mathematics, is what needs the next paper.
References
- Seamless Website Fingerprinting in Multiple Environments
- Encrypted internet traffic classification using a supervised spiking neural network
- Zero-shot temporal resolution domain adaptation for spiking neural networks
- DA-LIF: Dual Adaptive Leaky Integrate-and-Fire Model for Deep Spiking Neural Networks
- Realistic Website Fingerprinting By Augmenting Network Traces
- Deep Fingerprinting: Undermining Website Fingerprinting Defenses with Deep Learning
- Few-Shot Website Fingerprinting Attack
- ET-BERT: A Contextualized Datagram Representation with Pre-training Transformers for Encrypted Traffic Classification
- A new learning method for inference accuracy, core occupation, and performance co-optimization on TrueNorth chip
- k-fingerprinting: A Robust Scalable Website Fingerprinting Technique