Event cameras are sensors in which every pixel independently reports a tiny "event" whenever its brightness changes, instead of recording full images. They are often described as privacy-friendly because they never record a normal picture. That claim has never been quantified. The same physics that makes these sensors noisy works in privacy's favor. In the dark, leakage currents in each pixel fire spurious events at random, an effect called background activity (E-MLB), and the pixel's response becomes slow and smeared (v2e). This randomness hides fine detail about the scene, but only for a while. An observer who watches the stream longer can average the noise away.
This note proposes measuring that erosion in the units differential privacy uses. Differential privacy is the standard formal guarantee that an output reveals little about any one sensitive input. The tool is concentrated privacy accounting, a family of methods that treats privacy loss as a random variable and tracks its exact distribution as independent noisy releases pile up. We would model each short time window of an event stream as a randomized release of the scene, with its randomness supplied by the sensor's own physics. We would then compose these windows with exact numerical accountants. The result is a curve showing privacy budget spent against seconds of observation, at a given light level. On top of that curve we would build a sensor-side "privacy floor": the least extra randomization needed, with the intrinsic noise credited, to hold a target guarantee over a fixed horizon.
Core Claim
The fresh, per-window randomness in an event sensor can be written down as a known randomized mechanism over the scene. This randomness comes from shot and leakage background events and from jitter in when each event fires. Once written down, the privacy loss accumulated over many windows can be computed tightly, because its distribution is a convolution of per-window distributions. Two consequences follow, and both are new. The first is a calibrated, falsifiable statement of how private a raw event stream is as a function of illumination and observation time. The second is a sanitization mechanism that injects less artificial noise than one which ignores the sensor's own noise, at the same guarantee. The saving should be largest exactly where event cameras are deployed for their strengths: at night and in low light.
The threat model differs from the one prior spiking-network work uses. DPSNN trains spiking networks with DP-SGD, the standard method of clipping each training example's gradient and adding Gaussian noise. It protects whether a recording was in the training set, and it treats the event data as clean. We protect what is in the scene against anyone who receives the stream. Here "adjacent" means two scenes that differ by a bounded log-brightness change inside a sensitive region, such as a face patch. This is closer to local or metric privacy than to training-set membership.
The research file suggests a tempting alternative: let sensor noise stand in for DP-SGD's gradient noise. We reject that framing. A recording's noise is drawn once, when it is captured, and is then reused at every training step. It therefore supplies no fresh randomness across iterations and cannot be credited toward a training-set guarantee. Fixed pattern noise, the "hot pixels" caused by manufacturing defects (PDB Unet), and per-pixel threshold mismatch are similar. They are deterministic for a given sensor and must be treated as known to the adversary. Only noise that is redrawn in every window counts toward privacy.
Research Program
The per-window mechanism. For pixel $j$ and window $w$, we model the event count of each polarity as $n_{j,w} \sim \mathrm{Poisson}(\lambda_{j,w}(s) + \beta_j(I))$. Here $s$ is the scene's log-brightness trajectory, $\lambda$ is the signal-driven event rate after threshold jitter, and $\beta_j(I)$ is the background-activity rate. That background rate rises as illumination $I$ falls.
For adjacent scenes $s, s'$ producing rates $\Lambda = \lambda+\beta$ and $\Lambda' = \lambda'+\beta$, the privacy loss of one observation is
$$ \mathcal{L}_{j,w} = n_{j,w}\ln\frac{\Lambda_{j,w}}{\Lambda'_{j,w}} - (\Lambda_{j,w} - \Lambda'_{j,w}), $$
which is a scaled Poisson random variable. The total loss over the pixels in the sensitive region and $W$ windows is the sum of these terms. Its distribution is therefore the convolution of the per-term distributions. Background activity enters through $\Lambda$ in the denominator: more noise means a smaller log-ratio and less leakage per window.
We would fit $\beta$ and the jitter parameters per illumination level from real recordings. E-MLB records the same scenes under neutral-density filters that pass 1/4, 1/16 and 1/64 of the light, which gives controlled light levels. We would cross-check the fits against v2e's physical noise model (v2e).
Dependence across windows. The pixel's refractory period and its slow analog response correlate neighboring windows. We would do two things about this. We would choose window lengths longer than the measured correlation time and test independence empirically. Where correlation remains, we would bound the loss by treating each pixel as a short-memory process and composing over blocks rather than single windows. Showing honestly where the independence assumption holds is part of the contribution. It also echoes the open question in the file about time-correlated subsampling (Efficient privacy loss accounting for subsampling and random allocation).
Accounting machinery. We would take this machinery from existing work rather than invent it. Connect the Dots builds a piecewise-linear discretized representation of the privacy loss distribution that is guaranteed to be pessimistic, and supplies matching optimistic bounds that bracket the true curve. We would use it for the main curves. The Saddle-Point Method in Differential Privacy evaluates the tail of a composed loss analytically, in constant time and at very small failure probabilities. We need this because a sensitive region contains hundreds of pixels observed over thousands of windows. That paper reports that grid-based PRV accountants fail below a failure probability of $\delta = 10^{-10}$. EW-Tune corrects central-limit approximations for skew and kurtosis, which matter when only a few terms are composed. We would use it in the short-observation regime. Computing Exact Guarantees for Differential Privacy, the Fourier accountant, computes the same convolution through fast Fourier transforms and serves as an independent numerical cross-check.
The privacy floor. Given a target $(\epsilon,\delta)$ and horizon $T$, the sensor or its driver would do two things. It would add synthetic background events as an extra Poisson process with rate $\gamma(I)$, and it would randomly thin events in the sensitive region. Accounting would choose $\gamma$ and the thinning rate, with the intrinsic $\beta(I)$ credited. Because added Poisson noise simply sums with intrinsic noise in the rate $\Lambda$, the mechanism fits the same accountant without any new theory. At night it may need little or no injection.
Utility. We would measure what downstream models lose under each sanitization. Models are trained and evaluated on the sanitized streams. Classification uses N-Caltech 101 and CIFAR10-DVS (CIFAR10-DVS), and night-time car detection uses MVSEC-NIGHTL21.
Baselines
The naive reference point is doing nothing: report the raw stream as private with no stated guarantee. We would quantify this with the same accountant and with empirical attacks, so that "event cameras are private" becomes either a number or a refutation.
The strongest competitor for sanitization is a conventional local-privacy mechanism that ignores the sensor. It converts the stream to a fixed time-binned event histogram (a voxel grid), clips it, adds Gaussian noise, and accounts with the same tight accountant. For this baseline we found no reported result on Lacuna in the event setting. Running it ourselves with an identical accountant, identical horizon $T$ and identical $(\epsilon,\delta)$ makes the comparison fair. Any difference then comes from the mechanism and from the credit for intrinsic noise, not from looser bookkeeping.
The key ablation is our own Poisson privacy floor run with the intrinsic credit removed: $\beta$ is set to zero in the accountant, with everything else unchanged. This isolates the proposal's central ingredient. All downstream models share the same architecture, data split, training budget and hyperparameter search size for every arm. As a non-private ceiling we use the v2e result of 87.85% accuracy on N-Caltech 101 with a ResNet34 pre-trained on synthetic events and fine-tuned on real ones.
Beating the strongest baseline means two things. At matched $(\epsilon,\delta,T)$ and in the two darkest filter conditions, our mechanism must give higher accuracy or detection rate. It must also do so with the gap established as statistically significant across five seeds. If the ablation closes most of that gap, the credited noise is not doing the work, and we would report the result as such.
Why not just…?
Why not just measure privacy empirically by training an attacker on the stream? Attacks give lower bounds only. A weak attacker makes a leaky stream look safe, which is exactly the false comfort we want to remove. We still run attacks as an audit. An attacker is trained to tell adjacent scenes apart from $W$ windows, and its achievable true- and false-positive rates must lie inside the trade-off curve our accountant certifies. If the attacker ever falls outside that curve, the noise model is wrong. That outcome is the main way the proposal can fail, and it is a test we report rather than hide.
Why not just add Gaussian noise to event frames and use an off-the-shelf accountant? This is the real threat, and it is our main baseline. The file's objections to frame conversion in general — frames throw away fine timing (E-MLB), and spatial pooling under noise erases temporal cues (DPSNN) — concern training-time privacy and do not settle this question. What settles it is the experiment above. If the Gaussian-on-frames mechanism matches ours at night, then crediting intrinsic noise is not worth its modeling burden, and we would say so. We expect it to lose in the dark, because there $\beta$ is large and our mechanism injects little. In daylight we expect a tie, because there $\beta$ is negligible.
Why not just use Rényi or Gaussian differential privacy in closed form? Both are simpler, but the per-window loss here is a skewed, discrete Poisson variable. Closed-form central-limit summaries are known to misestimate composed loss (The Saddle-Point Method in Differential Privacy), and moment-based bounds are loose (Computing Exact Guarantees for Differential Privacy). EW-Tune reports that a refined accountant reduced the noise multiplier by up to 5.6% relative to Rényi accounting. We would report the gap between closed-form and tight accounting directly, as $\epsilon$ at fixed $T$ and as the observation time each method allows before reaching a fixed $\epsilon$. If the gap is negligible for these mechanisms, the simpler method should be recommended.
Proposed Experiments
Calibration. We first fit the noise model per light level on E-MLB's filtered recordings. The model passes if held-out window counts match the fitted Poisson-plus-jitter law, in their distribution and in their correlation across windows.
Privacy curves. We compute curves of $\epsilon$ against observation time for raw streams across the three filter levels and daylight DSEC driving sequences (DSEC). The adjacency is a bounded log-brightness change in a face-sized region. Each curve comes with pessimistic and optimistic bounds from Connect the Dots, plus a saddle-point evaluation of the deep tail.
Decisive checks. Three checks decide the proposal. The first is that the attacker's empirical trade-off curve always lies within the certified one. The second is that the certified horizon grows with darkness in the direction the physics predicts; a flat curve would mean the intrinsic noise is irrelevant. The third is the matched-$\epsilon$ utility comparison against Gaussian-on-frames, together with the no-credit ablation, on N-Caltech 101, CIFAR10-DVS and MVSEC-NIGHTL21. A positive outcome is a measurable, light-dependent privacy budget for event cameras and a cheaper sanitizer at night. A negative outcome — the attacker beating the bound, or no gain over the ablation — would still settle whether the privacy folklore around event cameras has any quantitative basis.
References
- E-MLB: Multilevel Benchmark for Event-Based Camera Denoising
- v2e: From Video Frames to Realistic DVS Events
- DPSNN: A Differentially Private Spiking Neural Network with Temporal Enhanced Pooling
- PDB Unet: A Spatio Temporal Video Fixed Pattern Noise Removal Network
- Efficient privacy loss accounting for subsampling and random allocation
- Connect the Dots: Tighter Discrete Approximations of Privacy Loss Distributions
- The Saddle-Point Method in Differential Privacy
- EW-Tune: A Framework for Privately Fine-Tuning Large Language Models with Differential Privacy
- Computing Exact Guarantees for Differential Privacy.
- CIFAR10-DVS: An Event-Stream Dataset for Object Classification
- DSEC: A Stereo Event Camera Dataset for Driving Scenarios