Hospitals, transit operators and sensor networks protect time-series records in two separate ways. They encrypt the records in transit, and they "date-shift" the records before release, moving every timestamp by a secret random offset so that no one can match the record to real-world events. The post-quantum problem makes the second layer matter more than it used to. Traffic encrypted today with RSA or elliptic-curve cryptography can be recorded now and decrypted later, once a large enough quantum computer exists. Anything that must stay confidential for decades, such as a medical record, therefore has to rely on de-identification alone.
This proposal asks how strong that last layer actually is. Date shifting leaves traces. Clocks jump by one hour for daylight saving time, clinics close on holidays, markets follow trading calendars, and notes contradict tables. We hypothesize that these traces let an attacker recover much of the hidden offset, and possibly the true year. We will build an attack that estimates the secret offset from such traces, measure the protection that remains in bits, and test a repaired shifting procedure that removes the leak without damaging the time intervals that analyses depend on.
Core Claim
Current releases treat a date shift as secure if the offset is drawn uniformly from a wide range. Two examples are the patient-level random shift in MIMIC-III, which covers 53,423 adult admissions and 7,870 neonates, and the season-preserving projection into the years 2150–2350 in MC-MED. Under that assumption, an attacker who wants to link records to outside logs such as check-ins, receipts or admission news faces about $\log_2 R$ bits of uncertainty, where $R$ is the number of possible offsets.
Our claim is that the artifacts shifting leaves behind lower this uncertainty, and that the drop can be measured. EHRCon, which built a benchmark for checking whether clinical notes agree with structured tables, found that over 56% of the disagreements in MIMIC-III involve time, many of them exactly one hour. Our hypothesis is that some of these one-hour gaps come from clock changes at daylight saving time. If so, they mark the true dates of the clock changes. Those dates follow fixed rules that differ between years, so their positions on the shifted calendar could narrow down the true year. Holidays, weekday closures and seasonality, which MC-MED deliberately keeps, would narrow it further.
We call the result the residual offset entropy. It is the uncertainty the attacker has left about the true offset $\delta$ after seeing the release $D$:
$$H_{\text{res}} = -\sum_{\delta} p(\delta \mid D)\,\log_2 p(\delta \mid D),$$
Here $p(\delta \mid D)$ is the attacker's posterior belief over offsets, computed from the calendar artifacts. We compare $H_{\text{res}}$ with the nominal $\log_2 R$. What is new here is treating de-identification as the component that carries long-term secrecy during the migration, and giving it a measured security level. Existing work either assumes the shift is secure or treats its artifacts only as a data-quality problem.
Relation to Prior Work
Two lines of work motivate the proposal.
The PQC Network Instrument study passively monitored 20.5 million OpenSSH connections over 16 months. Only 0.029% used post-quantum key exchange, and over 83% of servers dated from 2019 or earlier. Clients that support post-quantum schemes therefore fall back to classical ciphers. The study names medical records among the data exposed to later decryption, but it does not evaluate any application-layer defense. We supply one and measure it.
EHRCon documents time artifacts as a problem for checking consistency. Its best method, CheckEHR, reaches 61.06% recall in the few-shot setting. EHRCon never asks whether these artifacts help an attacker. We reuse its annotations to find artifacts, but our goal is recovering the offset, not detecting the disagreement.
Three lines of work supply the shifting methods we audit and attempt to repair.
- Static shifts. These are the patient-level shift in MIMIC-III and the season-preserving projection in MC-MED.
- Streaming temporal privacy. Beyond Value Perturbation defines ε-TLDP, a local differential privacy guarantee in which the protected change is swapping timestamps within a window of $k$ slots. It introduces threshold mechanisms that avoid empty or duplicated reporting slots.
- Stateful Switch. Stateful Switch uses atomic pairwise swaps of timestamps and loses no data.
The streaming guarantees cover the order of events within a window of $k$ slots. They say nothing about the absolute position of the whole stream on the calendar. Stateful Switch hides which minute a reading came from, not which year the stream came from. The two protections address different questions, and this proposal studies the one neither streaming paper covers.
One paper constrains the design. Beskar combines post-quantum key exchange (Kyber) and signatures (Dilithium) with value-domain differential privacy for federated learning. The value noise costs accuracy: at ε=10 on MNIST, local privacy gives 0.78 accuracy against 0.87 for central privacy. That cost is why we do not propose protecting timestamps by adding noise to values. Beskar also assumes the post-quantum channel succeeds, which the network study shows is often false today.
Research Program
Stage 1: the offset-recovery attacker
The attacker treats each candidate true offset as a hypothesis and scores how well the release fits the real-world calendar under that hypothesis. It uses four kinds of evidence:
- clusters of one-hour gaps between notes and tables, or within device streams, matched against historical daylight-saving rules;
- holiday drops in admission or trip volume;
- weekday structure;
- seasonal curves.
Each evidence source gives a likelihood, and the attacker combines them into the posterior $p(\delta \mid D)$.
Because the offset is shared across a patient, and in MC-MED the season is preserved, evidence can build up across many events for the same individual. For releases where every patient receives the same year window, evidence can also build up across the whole dataset.
Stage 2: measuring leakage with known ground truth
For streams where the true timestamps are known, we apply each shifting mechanism with a secret offset of our own and measure:
- the residual entropy $H_{\text{res}}$;
- top-1 offset recovery within ±1 day;
- the true-year recovery rate.
We also run a linkage attack. We generate a synthetic external footprint, for example check-in times for a random subset of individuals, and measure how often the attacker matches each footprint to the right record.
Stage 3: an artifact-aware shift
Given the leaks found in Stage 2, we build a shift that combines three repairs:
- Recompute clock-change hours on the shifted calendar, so that no one-hour gap sits at a true clock-change date.
- Restrict offsets to equivalence classes. Offsets are drawn only from sets of candidates that cannot be told apart by holidays or weekdays. This makes the nominal and residual entropy match.
- Add a Stateful Switch–style swap for fine-grained timestamps, so that artifacts at the minute scale are also hidden.
Throughout, we keep the within-record intervals $\Delta t$ exact. Preserving those intervals is what makes MIMIC-style shifting useful for survival models and trend analysis.
Baselines
Naive reference points. There are two bounds.
- The upper bound on leakage is releasing the data with no shift, where $H_{\text{res}}=0$.
- The "do nothing beyond the standard" reference is the nominal bound $\log_2 R$. This is what the MIMIC-III and MC-MED procedures implicitly claim.
Existing mechanisms. The strongest existing defenses are the season-preserving projection of MC-MED and the patient-level shift of MIMIC-III. They are the strongest competitors because they are what real releases use.
We rebuild both on the same ground-truth streams. The seed for the secret offset is held out from the attacker, and each mechanism gets the same range $R$. No reported result on Lacuna measures how recoverable the offset is under either mechanism, so their numbers will come from our own runs. For streaming data we also include Stateful Switch with a global offset added.
Utility. The utility baseline is the number reported for Stateful Switch: moving-average MSE 3 to 5 orders of magnitude lower than value perturbation, and more than 20% better utility than the RanSwitch baseline. On taxi-trajectory clustering, Beyond Value Perturbation reports NMI of 0.6–0.7, against near zero for Laplace value noise.
What counts as success. Our repaired shift beats the baselines if two conditions hold together:
- its residual entropy is within 1 bit of $\log_2 R$ while MC-MED-style projection falls clearly below;
- it matches their interval-based utility (moving-average MSE and clustering NMI) within noise.
Ablations. We apply each of the three repairs separately. This shows which artifact carries the leak, and whether the swapping step adds anything beyond the two calendar repairs.
Why not just…?
Why not just finish the post-quantum migration and skip de-identification? Deployment evidence says the gap lasts for years. After 16 months of measurement, adoption was 0.029%, and more than 83% of servers predated post-quantum support. Large post-quantum keys also overflow HTTP header limits in systems such as SciTokens (PQC Network Instrument). Traffic already harvested cannot be upgraded after the fact.
Honestly, the quantum argument is not required for our results. Credentialed public releases like MIMIC-III already face the same attacker without any decryption. We treat this as a strength: the audit matters for both threat models.
Why not just add Laplace noise to each timestamp? Independent noise would blur one-hour gaps, but it would also destroy the exact intervals between events. The evidence on value versus temporal perturbation (Stateful Switch, Beskar) shows that this kind of noise is costly. We include per-event noise on timestamps as an extra arm and report its utility on the same interval-based tasks. That lets readers judge the trade-off directly rather than trust our argument.
Why not just widen the shift range or scrub the one-hour bugs? This is the real threat to the proposal. It may be that cleaning up clock-change gaps alone brings the residual entropy back to nominal. Our first ablation tests exactly that. If it succeeds, the contribution shrinks to an audit plus a cleaning recommendation, and we will report it that way.
Widening the range is less promising. Clock-change rules and holiday calendars differ between years, so a wider range adds candidate years that the attacker can still tell apart. The seasonal and weekday structure that MC-MED keeps would also survive. We will sweep $R$ and plot $H_{\text{res}}$ against $\log_2 R$ to show whether the gap stays roughly constant as $R$ grows.
Proposed Experiments
Ground-truth leakage. On streams with true timestamps (the stock-price and taxi-trajectory feeds used in the temporal-privacy papers, and the Sepsis event log from PRIPEL), we apply each mechanism, inject realistic clock-change artifacts at the rate EHRCon observed, and measure $H_{\text{res}}$, true-year recovery and linkage success.
Consistency test without ground truth. On MIMIC-III and EHRCon we cannot know the true dates, so we run a falsifiable check instead. If the one-hour disagreements EHRCon annotates cluster near positions that, for a single constant offset per patient, correspond to clock-change dates of a real year inside 2001–2012, the leak exists in the released data. Offsets drawn at random should show no such clustering, which gives the null distribution for a permutation test.
Utility. We check utility on MC-MED's chronological 80/10/10 split and on moving averages and clustering over the streams.
Decisive check. The proposal succeeds if two results hold together: the attacker recovers the true year well above chance under the standard shifts, and the repaired shift removes that advantage while matching their utility.
References
- MIMIC-III, a freely accessible critical care database
- MC-MED, multimodal clinical monitoring in the emergency department
- EHRCon: Dataset for Checking Consistency between Unstructured Notes and Structured Tables in Electronic Health Records
- Post-Quantum Cryptography (PQC) Network Instrument: Measuring PQC Adoption Rates and Identifying Migration Pathways
- Beyond Value Perturbation: Local Differential Privacy in the Temporal Setting
- Stateful Switch: Optimized Time Series Release with Local Differential Privacy
- Efficient Full-Stack Private Federated Deep Learning With Post-Quantum Security
- PRIPEL: Privacy-Preserving Event Log Publishing Including Contextual Information