Explore the security challenges of heterogeneous computing systems in this 38-minute lecture by Hoda Naghibijouybari, Assistant Professor in the Department of Computer Science at Binghamton University. Delve into the evolution from traditional to modern heterogeneous systems, focusing on microarchitectural attacks and side-channel vulnerabilities. Examine GPU architecture, the CUDA programming model, and specific security threats like website fingerprinting and inter-keystroke timing attacks. Learn about defense mechanisms, including temporal and spatial partitioning, and the GPUGuard solution. Compare integrated and discrete GPUs, analyzing their unique security implications. Gain insights into the complexities of securing CPU-GPU systems, remote cross-component attacks, and potential defense strategies for heterogeneous computing environments.
Overview
Syllabus
Intro
Traditional Computing Systems
Modern Computing Systems Heterogeneous
Microarchitectural Attacks: Co-location
An Example: Side channel on Cache (Prime+Probe Attack)
GPUs are everywhere!
CPU vs. GPU
GPU Architecture
GPU Programming Model: CUDA
Covert Channel on GPUs?
Graphics/CUDA-Graphics Attack Overview
Attack1: Website Fingerprinting
GPU Memory Allocation Trace
Attack 2: Inter-keystroke timing
Security Issue
Temporal Partitioning
Spatial Partitioning
GPUGuard: Enable Intra-SM Sharing Securely
Integrated GPUs vs. Discrete GPUs
Security of Integrated CPU-GPU systems
Intel Integrated CPU-GPU
Main Challenges
Cache based covert channel
Remote Cross-Component Attacks
Defenses
Security of Heterogeneous Systems
Conclusion
Taught by
CAE in Cybersecurity Community
