Placing Private Needles in Private Haystacks: Factorized Synthesis for Rare-Event Detectors

Many detectors that matter most must find something that almost never happens. Examples include a bleed in surgical video that covers 0.8% of frames, a goal once in half an hour of football, or a stray abnormal heartbeat in a long recording. Training such detectors needs long recordings, and in medicine those recordings are exactly the data that cannot be shared. Differentially private synthetic data is the usual remedy: a generator trained with a formal guarantee that no single person's record changes its output much, which then releases artificial data for anyone to train on. Rare events are where this remedy is weakest. The noise that privacy requires falls hardest on the parts of the data that few records support. A generator trained on whole recordings also spends nearly all of its capacity on background that dominates 99% of the time.

We propose to split the generator in two. One private generator learns what background looks like. A second private generator learns only what short windows around events look like, and fills in the event so that it blends with the background around it. Synthetic training streams are then assembled by placing generated events into generated background at chosen positions. This gives exact timestamps for free and lets us choose how often events occur, independently of how good the generator is. The central question is whether this change in where the privacy budget and model capacity go turns synthetic rebalancing from harmful into helpful for rare-event detectors under a fixed privacy guarantee.

Core Claim

Two results in the file appear to conflict. Differential Privacy Under Class Imbalance studies imbalanced tabular data under privacy. It finds that generating a balanced private synthetic dataset (GEM) and then training an ordinary XGBoost classifier beats training classifiers privately, winning 7 of 8 metric rankings. It also proves that SMOTE-style interpolation is incompatible with privacy. Empirical-Distribution Matching for Synthetic ECG Classification studies a non-private ECG diffusion generator and finds the opposite for sequences. Rebalancing synthetic data toward rare classes hurts, and naively matching the real class frequencies gives the best macro AUROC of 0.8845, against 0.7827 for class prototypes. Its explanation is that the generator is worst where data is scarcest, so oversampling those regions multiplies artifacts.

Our claim is that this "rebalancing paradox" is mostly a property of monolithic generators, not of synthetic rebalancing in itself. If event windows get their own generator and their own share of the privacy budget, their fidelity should rise enough that moderate oversampling at assembly time improves downstream detection. We expect this effect to be larger under privacy noise than without it. The claim is falsifiable. If the factorized generator also shows the paradox, or if it does not beat a detector trained directly with private gradient descent, the idea fails.

Relation to Prior Work

Prior work plays three roles here: it motivates the gap, it supplies the method, or it constrains the design.

The two papers above motivate the work, because their conflict defines the question. Neither paper addresses temporal localization. The first works on static rows. The second classifies whole recordings and gives no formal privacy guarantee.

The private sequence generators supply method and also show the gap. GlucoSynth generates private glucose traces as 4-hour motifs and splits its privacy budget between a teacher-ensemble component and private gradient descent. We reuse the idea of splitting the budget across parts of the generator. We split by event versus background rather than by motif structure versus values, and we avoid its teacher partitioning, which relies on a 1.2-million-trace cohort. TIMEDIFF shows that diffusion models (generators that learn to turn noise into samples by gradual denoising) produce faithful clinical time series. However, it reports only empirical privacy risk and evaluates on mortality prediction, not event localization. HealthGAN bins streams into hours and finds that volatile subgroups come out "too consistent." This is the same failure we expect for rare events inside a monolithic generator.

Differentially Private Synthetic Heavy-tailed Data is closest in spirit. It shows that standard private clipping erases distribution tails unless the tails are modeled explicitly. Our event generator is a temporal analogue of giving the tail its own model.

Three papers constrain the design. Differentially Private Event Sequences over Infinite Streams protects sliding windows of aggregate counts. We need person-level protection for rich feature streams, so its window-level unit is too weak for us. Evaluating Differentially Private Synthetic Data Generation in High-Stakes Domains shows that fine-grained tasks degrade more than coarse ones: coreference F1 falls from 0.552 to 0.443 while simple classification falls only from 0.89 to 0.85. It also shows that demographic disparity widens, with the Equalized Odds gap growing from 0.20 to 0.30. This is a warning that temporal localization may be the hardest downstream task to serve with synthetic data.

On the detection side, RMS-Net and Mamba-OTR supply the downstream detectors. They need exact event offsets, and our assembly step supplies those offsets by construction.

Research Program

We work on per-frame feature streams rather than raw pixels or waveforms. Both target detectors already consume pre-extracted features, and private generation of full video is out of reach. The privacy unit is one whole recording, meaning one match, one video, or one patient. We train with person-level private gradient descent: sample recordings, clip each recording's total gradient, and add noise.

The background generator is a diffusion model over fixed-length feature windows drawn from anywhere in each recording. The event generator is a conditional diffusion model trained only on short windows centered on annotated events. It performs inpainting: given the background frames before and after a window from the background generator, it fills in the event span and its immediate surroundings. Each recording contributes at most $k$ event windows, which caps its influence. The two generators run on the same recordings, so the total guarantee comes from composing the two training runs. Under standard composition,

$$\epsilon_{\text{total}} = \epsilon_{\text{bg}} + \epsilon_{\text{ev}},$$

where $\epsilon_{\text{bg}}$ and $\epsilon_{\text{ev}}$ are the budgets spent on the background and event generators. We will use the tighter Rényi accounting in practice. The share $\rho = \epsilon_{\text{ev}}/\epsilon_{\text{total}}$ becomes an explicit design variable. Any choice of $\rho$ must be made without looking at private validation data, or the cost of choosing it must be paid from the budget.

The assembly step samples a background stream, chooses event positions according to a target rate $r$ and event-length statistics released privately, and inpaints events at those positions. The rate $r$ ranges from the empirical rate (no rebalancing) to several times higher.

Detectors such as RMS-Net or Mamba-OTR, with their usual focal or weighted losses, are trained non-privately on the released streams. Post-processing does not weaken the guarantee, so detector tuning on synthetic data is free.

The analysis has three parts. The first maps downstream accuracy as a function of $\epsilon_{\text{total}}$, $\rho$, and $r$. The second measures event-window fidelity separately from background fidelity, using a discriminator trained to tell real from synthetic event windows. The third checks whether detectors are exploiting assembly seams instead of real event features.

Baselines

The non-private ceiling is the published detector trained on real data. On SoccerNet, RMS-Net reports an Average-mAP of 65.5% with standard features and 75.1% with soccer fine-tuned features (RMS-Net). On EPIC-KITCHENS-100 take/release detection, Mamba-OTR reports 45.48 mp-mAP, and 43.35 mp-mAP under streaming evaluation (Mamba-OTR). We will retrain both on our feature pipeline so the ceiling matches our setup. A trivial "no event" predictor gives the floor.

The strongest private competitor that does not use synthesis is the same detector trained directly with person-level private gradient descent, using focal loss and event-centered cropping. Event-centered cropping is the trick that raised ECG F1 from 0.797 to 0.853 in Diagnosing Cardiac Abnormalities from 12-Lead Electrocardiograms. Cropping is applied inside the private sampler, so it costs nothing extra. This baseline gets the same $\epsilon_{\text{total}}$, the same compute, and the same hyperparameter-search budget, with that search accounted for in privacy. We found no reported result on Lacuna for privately trained detectors on SoccerNet or EPIC-KITCHENS. We will establish that number ourselves.

The strongest synthesis competitor is a single monolithic private diffusion generator over whole windows, with event labels as a conditioning channel. Synthetic data is sampled at the empirical event rate, the naive-bootstrap policy that won at 0.8845 macro AUROC without privacy (Empirical-Distribution Matching). We also run the same monolithic generator with conditional oversampling, which tests whether the paradox appears for monolithic generators in our setting. It has the same architecture family, total parameter count, total training compute, and $\epsilon_{\text{total}}$ as our method.

Two ablations isolate the key ingredient. The first is the factorized generator assembled at the empirical rate, which separates the effect of capacity and budget allocation from the effect of rebalancing. The second is the factorized generator with events pasted in without inpainting, which tests whether blending matters.

Our method beats the field if, at the same $\epsilon_{\text{total}}$ and with $\epsilon_{\text{total}} \le 8$, its best preregistered $r$ beats both private competitors on held-out real data. On SoccerNet the margin must be at least 3 Average-mAP points, the size of RMS-Net's own published gain. Results must be averaged over five seeds, with non-overlapping confidence intervals.

Why not just…?

Why not just train the detector directly with private gradient descent? This is the real threat. Detectors are smaller than generators, and per-recording clipping with event-centered crops may already keep the rare-event signal. The tabular evidence that private training trails a private generator plus a classifier (Differential Privacy Under Class Imbalance) may not transfer to sequences. The head-to-head comparison above settles it. If direct private training wins at every budget, the honest conclusion is that synthesis only buys reusability, meaning many detectors trained from one release, and we will report that.

Why not just use one conditional generator and sample more events from it? This is exactly what the rebalancing-paradox evidence warns against: oversampling where the generator is weakest multiplied artifacts and dropped macro AUROC to 0.7827 (Empirical-Distribution Matching). We run this design as a baseline rather than dismissing it. If it matches the factorized generator, our capacity-allocation explanation is wrong.

Why not just release synthetic data at the empirical rate and reweight the loss on it? Reweighting reuses the same few low-quality synthetic events with larger gradients. Fresh events at a higher rate give the detector new samples instead. The empirical-rate factorized ablation, with the detector's usual focal and weighted losses, measures exactly how much the extra rate adds beyond reweighting.

Why not just use window-level stream privacy? Window-level privacy as in Differentially Private Event Sequences over Infinite Streams protects any $w$ consecutive timestamps, not a whole patient or match, and it publishes counts rather than feature streams. It answers a weaker privacy question and cannot produce training data for a detector.

Proposed Experiments

The main testbeds are SoccerNet feature streams (500 matches, sparse goals and cards), with RMS-Net as the detector, and EPIC-KITCHENS-100 take/release detection, with Mamba-OTR under streaming evaluation. Both are public and serve as proxies where the privacy unit is a video. Their sizes, 500 matches and a comparable number of videos, are realistic for clinical cohorts. If access allows, we add the laparoscopic adverse-event set (Intraoperative Adverse Event Detection). It has 130 procedures with 0.8% bleeding frames and is the motivating case, but with so few units person-level privacy may only be achievable at large $\epsilon$, and we will report that plainly. As a direct test of the paradox at recording level, we repeat the PTB-XL sampling-policy comparison with the generator trained privately. This shows whether privacy noise alone changes the bootstrap-versus-rebalance ranking.

We sweep $\epsilon_{\text{total}} \in \{1, 4, 8, \infty\}$, $\rho \in \{0.2, 0.5, 0.8\}$, and $r$ from 1 to 8 times the empirical rate. Metrics are Average-mAP and mp-mAP on real held-out data, per-class average precision for the rarest classes, false-positive clusters per hour, and real-versus-synthetic discriminator accuracy measured separately on event and background windows.

The decisive checks are four. The first is the shape of accuracy against $r$. It should rise and then fall for the factorized generator, and fall monotonically for the monolithic one. If both curves fall, the claim fails. The second is whether the advantage widens as $\epsilon$ shrinks, which would mean factorization specifically counters privacy noise in rare regions. The third is a seam test. We train a probe to find insertion boundaries in synthetic streams, and we compare detectors trained on inpainted versus pasted events. A detector that has learned seams will show a large gap between synthetic validation accuracy and real test accuracy. The fourth is privacy auditing beyond the formal guarantee. We run membership inference at the recording level and nearest-neighbor matching of synthetic event windows to real ones, because Evaluating Differentially Private Synthetic Data Generation in High-Stakes Domains shows memorization can survive $\epsilon = 8$. Small event datasets are where such memorization is most likely.