Period-conscious Time-series Reconstruction under Local Differential Privacy
Yaxuan Wang, Tianxin Li, Enji Liang † \dagger, Yue Fu, Yanran Wang · 2026
Read Period-conscious Time-series Reconstruction under Local Differential Privacy by Yaxuan Wang, Tianxin Li, Enji Liang † \dagger, Yue Fu, Yanran Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-02724.txt.
Introduction
Yaxuan Wang and four co-authors present a method for rebuilding periodic time series after privacy noise has been added on the user's device. Periodic patterns drive many signals, from gait in video to rhythm in audio. Local differential privacy perturbs data before upload, which protects users, but the noise can corrupt spectral peaks and shift phase, wrecking period estimates.
Their framework first probes for the period at several scales and cross-checks candidates to suppress noise-made peaks. Samples are then grouped by matching phase within each cycle, stabilizing alignment across cycles. A denoising step that mixes an expectation-maximization method with density estimation recovers the per-phase values. On two real periodic datasets, the method kept periodic structure better and cut reconstruction error below standard privacy baselines, even under tight budgets.
Catalog record
| CREATOR | Yaxuan Wang, Tianxin Li, Enji Liang † \dagger, Yue Fu, Yanran Wang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.02724
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Period-conscious Time-series Reconstruction under Local Differential Privacy Yaxuan Wang, Tianxin Li, Enji Liang † \dagger, Yue Fu, Yanran Wang Abstract Periodic patterns are fundamental cues in multimedia signals and systems, including repetitive motion in video (e.g., gait cycles), rhythmic and pitch-related structure in audio, and recurring textures in image sequences. When such user-generated streams are collected from edge devices, local differential privacy (LDP) is appealing because it perturbs data before upload; however, the injected noise can corrupt spectral peaks and induce phase drift, making period estimation unreliable and degrading reconstruction quality. We propose CPR ( Cycle and Phase Recovery ), a period-aware reconstruction framework for periodic time series under LDP. CPR performs multi-scale period probing and multi-consensus selection to suppress noise-induced spectral interference, then aggregates perturbed samples at matched within-cycle phase positions to stabilize phase alignment across cycles. To recover the underlying per-phase values, CPR combines EM-based denoising with kernel density estimation, improving robustness under tight privacy budgets. Experiments on two real-world periodic datasets demonstrate that CPR better preserves periodic structure and consistently achieves lower reconstruction error than representative LDP baselines, especially in the low- ϵ \epsilon regime. I Introduction Periodic time series are ubiquitous in multimedia and multimodal sensing pipelines. Many multimedia tasks rely on stable periodic structure: repetitive motion patterns in video for action and gait analysis [ 19 ] [ 9 ] , pitch- and rhythm-related trajectories in audio for synthesis and enhancement [ 3 ] [ 1 ] , and recurring visual patterns across frames for texture analysis and reconstruction [ 16 ] . In practice, these periodic cues often appear not only in raw signals, but also in compact time-varying descriptors derived from multimedia content—e.g., pose/keypoint trajectories, motion energy, optical-flow statistics, pitch contours, or onset-strength envelopes—which are widely used for efficient indexing, summarization, synchronization, and downstream recognition. Beyond classic multimedia, similar periodic streams also arise in human-centric sensing (e.g., physiological monitoring) and other continuous modalities, where preserving cycle-level structure is critical for reliable modeling and decision making [ 7 , 10 ] . At the same time, collecting and sharing such streams raises serious privacy risks. Continuous video-, audio-, and sensor-derived time series can reveal identities, routines, and sensitive attributes (e.g., health conditions or behavioral patterns), even when the shared data are lower-dimensional features rather than raw media. Local Differential Privacy (LDP) provides a strong protection model by perturbing each user’s data before transmission, making privacy independent of the trustworthiness of the collector. Formally, for any two inputs t and t^{\prime} and any output t^{*} , \Pr[M(t)=t^{*}]\leq e^{\epsilon}\cdot\Pr[M(t^{\prime})=t^{*}], [ 18 , 17 , 5 , 4 ] where a smaller \epsilon enforces stronger privacy. This local model is particularly appealing for edge and mobile deployments with continuous data upload, where raw streams are hard to protect and auditing centralized processing is challenging. However, LDP noise can be especially destructive for periodic signals, particularly under strict budgets ( \epsilon<1 ) [ 17 , 8 ] . Because perturbation is injected at the sample level, it can (i) pollute the spectrum and blur dominant frequencies, making peak-based period estimation unstable; (ii) amplify apparent misalignment across cycles (often described as “phase drift”), where small errors in period detection or boundary localization accumulate into noticeable within-cycle offset; and (iii) exacerbate boundary artifacts in streaming publication, where noise accumulation and windowed budget splitting can break cross-cycle symmetry [ 11 ] . In multimedia settings, such failures translate into unstable cycle estimates, degraded reconstruction, and reduced utility for applications such as motion analysis, rhythmic modeling, or periodic texture reconstruction. Visually or perceptually, boundary effects can manifest as discontinuities between repeated segments—e.g., audible jitter in periodic audio or visible seams in reconstructed frame sequences—highlighting the need for reconstruction procedures that explicitly respect cycle/phase structure rather than only minimizing pointwise error. Existing LDP solutions do not adequately address this issue. Budget allocation strategies for streams (e.g., LBD [ 14 ] ) can improve efficiency, but they are not designed to robustly recover heterogeneous periodic structures common in real-world multimedia, such as variable period lengths, tempo drift, and multi-period superposition [ 13 , 2 ] . Other reconstruction approaches (e.g., compressed sensing based methods) may rely on assumptions that are difficult to guarantee in practical deployments [ 15 ] , limiting robustness for high-quality reconstruction. These gaps motivate a period-conscious view of LDP reconstruction: the goal is not only to denoise, but to stabilize the dominant cycle and align phases so that the reconstructed stream preserves the structure most relevant to periodic multimedia analysis. To address these challenges, we propose CPR ( Cycle and Phase Recovery ), a period-conscious reconstruction framework for periodic time series under LDP. Real-world signals can be complex with many rhythms. However, in strong LDP noise, small rhythms are usually completely covered by noise spikes. CPR focuses on finding the most stable period because it is the most useful part of the signal that can still be saved. This approach is practical and ensures the best possible recovery under high privacy protection. The main contribution of this paper is three-fold: • A period-conscious reconstruction problem formulation for LDP-protected multimedia streams. We highlight how LDP noise corrupts periodic structure through spectral pollution, phase drift, and boundary effects in continuous publishing, motivating reconstruction methods that explicitly preserve cycle/phase information. • Multi-scale, multi-consensus period identification under LDP. CPR probes candidate periods across multiple window sizes and selects reliable cycles via cross-window agreement, suppressing spurious peaks caused by privacy noise and supporting heterogeneous periodic patterns. • Phase-aware aggregation and density-based reconstruction. Given the estimated period, CPR groups perturbed observations by within-cycle phase, stabilizes cross-cycle alignment, and reconstructs per-phase values using an EM-then-KDE strategy, yielding improved accuracy and robustness under tight privacy budgets. II Preliminaries II-A Notation and normalization We consider a length- n time series X^{\mathrm{raw}}=\{x^{\mathrm{raw}}_{t}\}_{t=1}^{n} . Each device maps values to a bounded domain for privacy mechanisms defined on [0,1] : x_{t}=\mathrm{Normalize}(x^{\mathrm{raw}}_{t},[0,1])\in[0,1],\qquad X=\{x_{t}\}_{t=1}^{n}. (1) II-B Approximate periodicity and dominant period A standard proxy for “how periodic” a series is at lag T is the mean squared shift discrepancy \Delta_{T}(X)=\frac{1}{n-T}\sum_{t=1}^{n-T}(x_{t}-x_{t+T})^{2}. (2) A dominant period is a T (within an admissible range) for which \Delta_{T}(X) is comparatively small. II-C w -event local differential privacy and budget splitting We adopt w -event LDP for streams [ 14 ] . Let I=\{t,\dots,t+w-1\} be any contiguous window of length w , and let M_{I} be the joint release over I . A mechanism satisfies \epsilon - w -event LDP if for any two window inputs u,v\in[0,1]^{w} and any measurable output set \mathcal{O} , \Pr[M_{I}(u)\in\mathcal{O}]\leq e^{\epsilon}\Pr[M_{I}(v)\in\mathcal{O}]. (3) If each per-time mechanism is \epsilon_{0} -LDP, then by sequential composition any length- w window is (w\epsilon_{0}) -LDP. Hence we set \epsilon_{0}=\epsilon/w, (4) so the stream satisfies ( 3 ). II-D Square Wave (SW) local randomizer We use the SW mechanism [ 12 ] as the per-time local randomizer. Given x\in[0,1] , SW outputs y\in[0,1] with a two-level conditional density: f_{\mathrm{SW}}(y\mid x)=\begin{cases}p,&y\in I_{x},\\ q,&y\in[0,1]\setminus I_{x},\end{cases} (5) where I_{x} is a length- 2b interval “centered” at x but shifted to stay inside [0,1] : I_{x}=\begin{cases}[0,2b],&x<b,\\ [x-b,x+b],&b\leq x\leq 1-b,\\ [1-2b,1],&x>1-b.\end{cases} Parameters (b,p,q) are determined by \epsilon_{0} (with p/q=e^{\epsilon_{0}} ) as in [ 12 ] : b=\frac{\epsilon_{0}e^{\epsilon_{0}}-e^{\epsilon_{0}}+1}{2e^{\epsilon_{0}}\bigl(e^{\epsilon_{0}}-1-\epsilon_{0}\bigr)},\qquad p=\frac{e^{\epsilon_{0}}}{2be^{\epsilon_{0}}+1},\qquad q=\frac{1}{2be^{\epsilon_{0}}+1}. Applying SW independently at each time