Collaborative Research: CNS Core: Small: Privacy by Memory Design

Project Synopsis

      Differential privacy (DP) has been widely accepted as the de facto technique for protecting data privacy. Despite the decade-long research efforts on DP, there still exists a critical research problem that has been largely overlooked, that is all existing DP studies are grounded on the hypothesis that software can easily and faithfully sample and add noises from a probability distribution. However, this hypothesis is being constantly challenged by recent findings about its privacy violation and by the growing demand of privacy protection in low-end devices that may lack high-level software libraries. Hence, this project's innovative research angle is to realize DP mechanisms directly on embedded memories, which are ubiquitous in modern electronic devices.

      The fundation of this project is based on the physical property of embedded memeories - data bits stored in memory cells will be corrupted when the supply voltage of the memory is reduced below the nominal voltage. As shown in Figure 1, the cell failure rate is dependent on transistor sizes, supply voltages and several other parameters. This project aims to design and optimize memory architecture and peripheral circuits to control cell failures to let stored bits be corrupted in a way that conforms to the DP notions. Towards this goal, this project proposes three research thrusts:

      (1): Design Memory-based (Central and Local) DP Noises and Processing Algorithms;

      (2): Design and Optimize Memories for Flexible and Robust (Central and Local) DP;

      (3): Chip Tape-out, System Implementation, Evaluation, and Case Studies.

      In our prior studies [TIFS'22], we developed a prototype to demonstrate the realization of Local DP (LDP) in SRAM memory and the achievement of significant power savings on memory read/write processes. Our idea was based on a key observation that memory cells exhibit failures at sub-nominal supply voltages, which perturbs data bits (e.g., 0 to 1) stored therein. To utilize this characteristic towards the LDP realization, we designed a new memory architecture with a hybrid 6T-8T cell structure that achieves heterogeneous cell failures when operating at sub-nominal supply voltages. With such design, when data is read out from the memory, some bit positions can be controlled to flip their bit values while others can have their bit values remain unaltered. The cell failures (or bit flipping) are properly controlled via runtime power control to realize the randomized response on some specific binary bits of the stored data, thereby ensuring LDP on the original data. In the meantime, we purposefully rendered cell failures to the least significant bits to minimize the utility loss.

Personnel and Collaborators

  • PhD students
  • Collaborators
  • Dr. Na Gong (site-PI)
    W. Nicholson Associate Professor
    Department of Electrical and Computer Engineering
    University of South Alabama
    Email: nagong at southalabama dot edu

    Research Progress

  • Publications
      [1] Two Birds with One Stone: Differential Privacy by Low-power SRAM Memory
              J. Liu*, N. Gong*, and H. Das, IEEE Transactions on Information Forensics and Security (IEEE TIFS), Under Review. (*equal contribution)

      [2] Query Integrity Meets Blockchain: A Privacy-Preserving Verification Framework for Outsourced Encrypted Data
              S. Jiang, J. Liu, Y. Liu, J. Chen, L. Wang, and Y. Zhou, to appear at IEEE Transactions on Services Computing (IEEE TSC)

  • Education and Outreach Activities

    under development


    Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.