Padding Clinical AI Traffic Within the Bedside Action Window

A hospital machine-learning model that predicts sepsis or a dangerous lab result does not only produce a prediction. It also produces network traffic. The model service is queried when a clinician signs an order. It answers with a risk score. An alert is pushed to a workstation, and someone opens the chart. All of this is encrypted, but encryption hides only the content of messages. Packet sizes, timing and bursts stay visible to anyone on the path. In smart homes, that visible pattern alone lets an eavesdropper infer user activity with nearly 100% accuracy (Keeping the Smart Home Private with Smart(er) IoT Traffic Shaping). Real-time clinical deployments increasingly route inference through public cloud functions that talk to the hospital's electronic medical record, as in DEPLOYR. So the same leak could reveal that a patient on a given unit is deteriorating, without a single byte being decrypted. The standard fix is traffic padding: send fake "cover" traffic and delay real traffic so the pattern looks the same whatever is happening. Padding costs bandwidth and adds delay, and delay is exactly what clinicians are assumed not to tolerate.

This proposal argues that the tolerable delay for a clinical alert is set by the clinical action loop, not by network latency. That loop is the time it takes a nurse or physician to go to the bedside, confirm the condition and act. Frontline clinicians using a deployed sepsis model said that alerts arriving before symptoms can be checked at the bedside are distrusted and ignored (Human-machine teaming is key to AI adoption). If that holds, many alerts carry a hidden slack window. Within that window, real traffic can be held and released on a fixed schedule at almost no clinical cost. Only alerts with tight windows need costlier, randomized cover traffic. The windows are not written down anywhere. We propose to measure them per alert class with semi-structured interviews, which are open-ended conversations that follow a common guide. We then use the measured windows as the delay budgets of a padding scheme, and test on replayed ICU data whether this beats generic padding on privacy, bandwidth and missed-deadline rate.

Core Claim

A traffic shaper for clinical inference that uses a separate, empirically measured delay budget for each alert class can push an attacker's inference of patient-state events to chance level. It should do so with lower bandwidth overhead than stochastic traffic padding tuned to a single uniform delay, and with no more deadline violations. The new ingredient is the budget, not the padding mechanism. Prior padding work either fixes delay globally or optimizes it against an abstract cost. Prior clinical qualitative work never turned its findings into system parameters. The claim fails in three cases, and the experiments are built to detect each. The first is that clinical traffic does not leak much to begin with. The second is that a plain fixed schedule with fixed-size messages is already cheap enough. The third is that budgets from a quick survey or from audit-log timing match interview-derived budgets.

Research Program

Testbed and threat model. No clinical traffic trace dataset exists in the material, so we build one. We use PyHealth 2.0, a toolkit that turns MIMIC-IV records into patient event streams, to replay timestamped ICU events: lab results, orders and vital-sign updates. The replay drives a DEPLOYR-style architecture. A mock medical-record server triggers cloud-hosted models in two ways, on scheduled cycles (DEPLOYR uses 15-minute ICU triggers) and on events such as an order being signed. Responses then flow back as alerts and chart opens. All traffic crosses a TLS link that we capture at the hospital–cloud boundary. That boundary is where an upstream provider or a network tap sits.

Labels come from the replayed record: sepsis-onset windows, abnormal-lab windows and rapid-deterioration windows. The attacker's task is to say, from a short window of encrypted traffic, whether such an event is occurring. The attacker is the one-dimensional convolutional network from Deep Fingerprinting. That network reached 98.3% accuracy on undefended Tor traffic and still 90.7% against the adaptive-padding defense WTF-PAD. We retrain it on each defense's own traces, so every defense faces an adapted attacker.

Interview study. We will run semi-structured interviews with roughly 25 participants: ICU and ward nurses, physicians, rapid-response staff and clinical informatics or network staff. This sample size is in line with Henry et al. (20 clinicians) and the 17 interviews in the pediatric mixed-methods study. Following the pre-conditioned design of Masden et al., each interview starts from a short survey of the participant's role and the alert types they handle. The guide is then tailored to those alert types.

For each alert class, the guide walks through recent concrete cases. It asks when the participant could first have acted, what they had to verify first, and how much later the alert could have arrived without changing what they did. Its purpose is to surface informal rules nobody has documented, the same thing such interviews recovered in the Australian SME study. A second strand asks informatics staff about network paths and logging. Its purpose is to decide where cover traffic must be dropped so that it never creates a phantom record or an alert. That matters because spurious alerts feed both distrust of AI and alarm fatigue (Gaube et al.).

Two coders will code transcripts independently to saturation, as in Murray et al.. The output is a budget $B_c$ for each alert class $c$: a lower-bound estimate of the extra delay that would not change clinical action, with uncertainty taken from the spread across participants.

Budget-aware shaper. The shaper runs on a fixed slot schedule with fixed-size messages, and real messages ride in the slots. Suppose a real message of class $c$ arrives at time $t$, and the next scheduled slot is at $s(t)$. The message is deferred to that slot if

$$s(t) - t \le B_c,$$

and otherwise it is sent immediately. For classes whose budget is shorter than the slot spacing, immediate sends are hidden by randomized cover bursts. These follow the stochastic traffic padding of Smart(er) IoT Traffic Shaping, which matches a profiled burst rate and duration and injects identical fake bursts with probability $q$. Cover traffic ends at a gateway before the medical-record server. We tune the slot spacing and $q$ jointly to minimize bandwidth, subject to two conditions: attacker accuracy within a tolerance of chance, and zero deadline violations on held-out replay.

Baselines

The naive reference point is no defense: the testbed traffic sent as is. This shows how much clinical inference traffic actually leaks. If the adapted attacker cannot beat chance here, the project's motivation is gone, and we will report that.

The heavy-handed reference is constant-rate, constant-size link padding, which sends a message every slot whether or not there is real data. Its cost in smart homes was over 100GB of cover traffic per month. Stochastic padding reached similar protection with 1.07x to 1.68x overhead (Smart(er) IoT Traffic Shaping). Clinical model payloads are small, so it may be far cheaper here. It is run at several slot spacings: at short spacings it meets every budget, and at longer spacings it costs less bandwidth but misses more deadlines.

The strongest existing competitor is stochastic traffic padding itself, with one uniform delay bound chosen to meet the tightest alert class. It is the best published mechanism for low overhead on bursty device traffic, and it already reduced adversary confidence to guessing level in the IoT testbed. We also run the phase-specific interval padding of PriVA-C. PriVA-C cut voice-command fingerprinting from 89% to 13% at 156% overhead, versus 1003% for the constant-rate scheme BuFLO, and it is the closest prior work that tied padding parameters to human tolerance. We also run the event-level differential-privacy shaper of Network Traffic Shaping for Enhancing Privacy in IoT Systems. It offers a formal guarantee and is swept over privacy levels. No reported result was found on Lacuna for any of these defenses on clinical inference traffic, so all are re-run on our traces.

Fairness rules apply to every defense. Each defense gets the same traces and the same attacker retrained on its own output. Each also gets the same search budget over its parameters, and is required to satisfy the same per-class deadlines on a validation split. Beating them means being dominant on the trade-off curve: equal or lower attacker accuracy at strictly lower bandwidth overhead, with equal or fewer deadline violations, on held-out patients.

Two ablations isolate the key ingredient. The first keeps our shaper but sets every $B_c$ to the tightest elicited budget, which removes the per-class structure. The second replaces interview-derived budgets with budgets from a short Likert survey of the same participants, plus timing statistics mined from MIMIC-IV order and result timestamps.

Why not just…?

Why not just send fixed-size messages on a fixed schedule? This is the most serious threat. DEPLOYR-style ICU models already run every 15 minutes, and the payloads are small. Plain constant-rate padding may therefore cost little, unlike the video-heavy smart-home case. The constant-rate baseline is designed to settle this. If constant-rate padding meets every deadline at overhead comparable to ours, the right conclusion is that clinical inference does not need stochastic padding. We would report that the contribution reduces to the measured budgets that justify the schedule. Our expectation of a gap rests on event-triggered and alert-delivery traffic. That traffic is either delayed beyond some budgets by a long schedule, or made expensive by a short one. But that is an empirical question.

Why not just keep inference on-premise, or rely on the VPN? VPNs hide where traffic goes, not how much there is or when it is sent (Smart(er) IoT Traffic Shaping). Deployments like DEPLOYR deliberately cross a cloud boundary. On-premise hosting removes the upstream-provider adversary but not an insider tap on the hospital network. We will repeat the attack with the capture point moved inside the hospital network. That shows how much leakage survives a fully on-premise design.

Why not just get delay budgets from a survey or from audit-log timing instead of interviews? Logs record when clinicians acted, not when they could have acted or what they had to verify first. A survey number has no reasoning behind it, and the pediatric study needed its interviews to explain the survey rankings (Perspective Toward Machine Learning Implementation in Pediatric Medicine). Still, this is testable rather than assumed. The second ablation directly compares shapers built from survey-plus-log budgets with shapers built from interview budgets. If they perform the same, the cheaper elicitation wins and we will say so.

Proposed Experiments

The first decisive check is leakage. On undefended traces, does the adapted attacker detect sepsis-onset and deterioration windows well above chance? We report balanced accuracy and AUROC per event type.

The second check is the trade-off curve. For every defense, we plot attacker balanced accuracy against bandwidth overhead and the fraction of alerts delivered past their class budget, all on held-out patients. The claim holds only if the budget-aware shaper dominates both stochastic padding with a uniform delay and constant-rate padding.

The third check is the two ablations. They show whether the gain comes from per-class budgets and whether interviews add anything over survey and log timing.

The fourth check is budget validity. We present a separate group of clinicians with replayed case vignettes in which alerts are delayed by randomized amounts within and beyond the elicited budgets. We then run short follow-up interviews about whether their decisions or trust would have changed. This catches budgets that participants overstated. It tests the Henry et al. observation that early alerts are not actionable.

Finally, we audit the mock medical-record logs and alert queue over the full replay. The requirement is that no cover message ever produces a record or an alert.