EpiOracle: Privacy-Preserving Cross-Facility Early Warning for Unknown Epidemics

Shiyu Li, Yuan Zhang, Yaqing Song, Fan Wu, Feng Lyu, Kan Yang, Qiang TangProc. Priv. Enhancing Technol 2025 · DOI

Introduction

Traditional epidemic early warning systems are primarily disease-based, relying on confirmed diagnoses. However, the emergence of unknown pathogens—often referred to as "Disease X"—renders these systems ineffective during the critical initial stages of an outbreak. Syndrome-based surveillance, which monitors symptoms like fever or cough, offers a more proactive alternative. Yet, existing syndrome-based systems often require healthcare facilities to upload plaintext data to central servers, posing significant risks of sensitive data leakage and "frequency leakage," where the statistical prevalence of symptoms could be exploited for economic gain or to incite public panic.

The EpiOracle Framework

EpiOracle is a privacy-preserving framework designed to detect unknown epidemic outbreaks by analyzing syndrome patterns across multiple healthcare facilities without exposing raw data. Unlike systems that rely on computationally expensive Fully Homomorphic Encryption (FHE), EpiOracle utilizes a highly efficient "Extract-then-Encrypt" approach based on standard symmetric-key primitives like AES and SHA-3.

Extract-then-Encrypt Mechanism

The core of EpiOracle's fuzzy detection lies in its ability to group similar symptom lists. Patients with the same underlying disease rarely present identical symptoms. To address this, the system extracts features from symptom lists to derive encryption keys.

The Extract-then-Encrypt framework for fuzzy symptom matching.

As shown in the diagram, if two facilities possess similar symptom lists ($m_1$ and $m_2$), they can derive the same cryptographic key. This key allows them to decrypt a "tag"—a random string representing a specific syndrome group—facilitating cross-facility matching without revealing the underlying symptoms to the central cloud.

Oblivious Counting and Permutation Matching

To track syndrome frequency while maintaining confidentiality, EpiOracle employs a novel oblivious counting mechanism based on a variant of Bloom Filters. When a match is found, the cloud flips a random bit in a specific segment. The cloud observes bit-flipping activity but remains unaware of which syndrome is being counted or the exact count until a predefined threshold triggers an alert. Additionally, the system uses permutation matching to ensure accuracy even when symptoms are reported in different orders.

Evaluation and Results

The researchers evaluated EpiOracle using real-world COVID-19 symptom data, synthesized WHO-based datasets, and noisy background data to test the system's ability to distinguish between background noise and potential outbreaks.

  • Accuracy: The system achieved approximately 70% accuracy in detecting similar symptom lists from real-world COVID-19 cases.
  • Efficiency: EpiOracle demonstrates remarkable performance, with a facility-side computation delay of only 12.4 ms, making it practical for real-world deployment.
  • Communication: The communication overhead is minimal, requiring less than 40 KB for every 100 similarity checks, which is feasible even for facilities with constrained bandwidth.

Conclusion

EpiOracle represents a significant step forward in public health surveillance by prioritizing both data content and frequency confidentiality. By favoring efficient symmetric-key algorithms over complex Multi-Party Computation (MPC) solutions, it offers a practical, hardware-independent path toward global early warning systems for future pandemics. Future work will focus on refining similarity metrics beyond Hamming distance and further mitigating potential facility-side data leakage.

Summary updated

BibTeX
@article{li2025epioracle,
  author = {Shiyu Li and Yuan Zhang and Yaqing Song and Fan Wu and Feng Lyu and Kan Yang and Qiang Tang},
  title = {{EpiOracle: Privacy-Preserving Cross-Facility Early Warning for Unknown Epidemics}},
  journal = {Proc. Priv. Enhancing Technol.},
  year = {2025},
  doi = {10.56553/popets-2025-0020},
}
Figure