Yoshi Sato

This thesis presents an architectural and empirical characterization of Coconut-SVSM, an open-source paravisor for AMD SEV-SNP confidential virtual machines. Through source-level analysis and experiments on a physical AMD EPYC server, it examines Coconut-SVSM's architecture, vTPM integration, protocol-call latency, memory overhead, and boot performance. The results show that Coconut-SVSM can provide useful in-guest vTPM functionality with low steady-state protocol and memory overhead, while identifying practical limitations in startup, persistent state, and remote-attestation integration.

Modern software security increasingly relies on bug bounty pro- grams, which incentivize independent researchers to discover and responsibly disclose vulnerabilities. While effective in traditional software ecosystems, these programs face unique challenges in the context of smart contracts. Smart contracts run on blockchains, manage financial assets, and are typically authored by pseudony- mous developers. As a result, when vulnerabilities are discovered, security researchers often lack a secure and reliable channel for disclosure. Existing workarounds, such as encrypting messages to contract authors' public keys, introduce inefficiencies and privacy risks, since ciphertexts themselves may reveal the presence of bugs. We propose DeadDrop, the first system to support oblivious bug reporting for smart contracts. DeadDrop combines oblivious message retrieval (OMR) with a trusted execution environment (TEE) to enable researchers to privately and efficiently deliver vul- nerability reports without revealing their intended recipient or overwhelming authors with spam. Our design specifies security requirements for such a system, presents practical techniques for handling long messages, and introduces a bug specification lan- guage to formalize submissions. We implement a prototype and evaluate its performance, demonstrating that oblivious bug report- ing is both feasible and efficient, achieving an amortized processing time of approximately 3 ms per submission. Finally, we discuss in- centive mechanisms to encourage participation, highlighting open challenges for decentralized bug bounty ecosystems.

Confidential Virtual Machines (CVMs) like Intel TDX and AMD SEV-SNP provide hardware-based isolation for cloud workloads. While SEV-SNP supports GPU passthrough for legacy (non-CC capable) accelerators, Intel TDX currently lacks this capability. This work-in-progress analyzes the performance implications of enabling such functionality across both platforms. We identify fundamental architectural differences: SEV-SNP’s guest-controlled page encryption enables near-native performance, while TDX’s hypervisor-mediated design requires mandatory VMEXITs and may impose significantly higher costs. Our CUDA microbenchmarks establish baseline metrics showing pageable transfers achieve 50-75% throughput of pinned DMA on Intel Xeon systems. We present the technical requirements for TDX GPU passthrough and analyze why Intel’s stricter security model complicates performance. This work provides developers with critical performance expectations when choosing between CVM platforms for GPU workloads, highlighting the trade-offs between security architectures and practical performance.

CoGIA is an NSF-funded project developing a privacy-preserving platform for genome imputation and analytics. The project combines trusted execution environments with data-oblivious computation, security monitoring, and differential privacy to protect both users' genomic data and controlled-access reference panels. Its broader goal is to enable deployable confidential genomic services while developing foundational techniques for secure computation over sensitive biomedical data.