Event
Ph.D. Research Proposal Exam: Olsan Ozbay
Wednesday, October 21, 2026
3:00 p.m.
AVW 2328
Souad Nejjar
301 405 8135
snejjar@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Olsan Ozbay
Committee:
Professor Ankur Srivastava (Chair)
Professor Dana Dachman-Soled
Professor Manoj Franklin
Date/time: Wednesday, October 21 at 3:00 PM
Location: AVW 2328
Title: Security Techniques for Heterogeneous Integration
Abstract: Heterogeneous integration combines separately manufactured components into a higher-level assembly.
In the mixed-trust 2.5D/3D systems considered here, sensitive information can be exposed in two contexts: computation involving private data on untrusted chiplets, and verification exchanges between system integrators and chiplet vendors.
This proposal characterizes the information exposed in these contexts and develops approaches for limiting that exposure.
The first part of this proposal addresses the confidentiality of private inputs, intermediate values, and outputs during computation on untrusted chiplets.
Without a protection mechanism, offloading computation exposes private operands and results in plaintext to malicious logic on an untrusted chiplet.
We therefore develop two information-theoretic approaches to address this problem.
The first approach, Private Information Retrieval for Directed Acyclic Graphs (PIR-DAG), decomposes an n-input Boolean function into a directed acyclic graph of w-input look-up tables and applies private information retrieval (PIR) to each table.
This replaces retrieval from a single 2^n-entry truth table with a sequence of retrievals from smaller 2^q-entry tables and ensures that, provided the k untrusted chiplets do not collude, no individual chiplet learns information about the private inputs, intermediate values, or outputs.
Because physical co-packaging alone does not enforce non-collusion, the second approach, Broadcast PIR (BPIR), removes this assumption by reversing the dataflow: the untrusted chiplets broadcast public look-up tables, while all input-dependent selection remains within the trusted chiplet.
The second part of this proposal addresses disclosure during chiplet verification.
We show that automatic test pattern generation (ATPG) reports can act as structural fingerprints of a vendor's netlist, while coverage requests can expose the integrator's design intent.
This leakage can arise under fully compliant, honest-but-curious behavior, so neither physical inspection nor authentication addresses it.
We propose Digital Twin mediation as a disclosure-control point and formulate the resulting privacy--utility tradeoff as the central design problem.
