In this episode of the Computer Architecture Podcast, hosts Dr. Suvinay Subramanian and Dr. Lisa Hsu welcome Professor Todd Austin from the University of Michigan. Professor Austin, a distinguished figure in electrical engineering and computer science, is known for his extensive work in robust and secure system design, hardware and software verification, and performance analysis. With a rich background that includes roles as a senior processor architect at Intel's microprocessor research labs, a professor at the University of Michigan, director of research centers like Cfar, and CEO and co-founder of Agita Labs, Professor Austin brings a wealth of experience to the discussion. He is also an IEEE fellow and a recipient of the ACM Morris Wilkes award for his contributions to SimpleScalar and the Diva and Razor architectures.
The central theme of this episode revolves around "durable security" and privacy-enhanced computing. Professor Austin shares his insights on the challenges of building systems with more robust and lasting security than current solutions. He delves into the complexities of eliminating zero-day bugs and vulnerabilities, drawing from his experiences working with DARPA and his current work at his startup, Agita Labs, focusing on building durable security systems, particularly within the Azure cloud. The conversation explores the nuances of non-durable security, exemplified by practices like "Patch Tuesday," and the inherent difficulties in creating truly unhackable software and hardware.
Chapters
00:00:13 — Introduction to Professor Todd Austin
00:01:31 — What gets Todd up in the morning: durable security
00:02:21 — Defining durable security: Patch Tuesday, SGX, and side channels
00:05:54 — Why side channels are inherent to high-performance design
00:06:50 — Kocher's RSA attack and the Leaky Frontends paper
00:09:50 — Cryptography and isolation as durable defenses (Open Titan)
00:11:35 — Morpheus: encrypting pointers and surviving DARPA red-teaming
00:13:45 — From encrypting pointers to computing on encrypted data
00:15:08 — From security to privacy: the transformative shift
00:16:55 — Privacy-oriented architectures and the voting-machine thought experiment
00:19:06 — The hardest problem: safely decrypting selected results
00:22:12 — Three approaches to encrypted compute: FHE, MPC, and hardware encryption
00:24:55 — High-entropy ciphertext and a redefined threat model
00:28:53 — Overheads: 50× off native today, within 2× in simulation
00:30:19 — Data-oblivious algorithms: predication, ORAM, and the floodfill case study
00:34:32 — Reducing adoption friction: libraries, frameworks, and automation
00:38:27 — Career arc and capacity-building in Ethiopia
00:41:18 — Origin story: Wisconsin EE, Xerox, and the path to a PhD
00:46:09 — Postdoc at Intel and the Diva project at Michigan
00:48:08 — A rule-breaking philosophy for research
00:49:32 — What the architecture community does well and where it falls short
00:50:35 — Open-source hardware, RISC-V, and the cost of innovation
00:53:52 — Design agility and the future of hardware description languages
00:57:27 — FPGAs everywhere: an exciting future for hardware
00:59:58 — Closing thoughts and farewell
Takeaways
Durable Security is Proactive, Not Reactive: True durable security aims to build systems that inherently resist attacks and vulnerabilities, rather than relying on constant patching and fixes after exploits are discovered.
Side Channels are a Fundamental Challenge: Eliminating side channels, where information leaks through shared hardware resources or timing variations, is a critical and difficult problem in achieving robust security.
Privacy-Enhanced Computing Enables Secure Data Collaboration: Technologies that allow computation directly on encrypted data can revolutionize how sensitive information is processed, enabling value extraction without compromising raw data privacy.
Open Source Hardware is Key to Innovation: Adopting open-source principles for hardware (like RISC-V) can significantly lower the barrier to entry and cost of innovation, allowing resources to be focused on value-added differentiation rather than re-creating common infrastructure.
Algorithmic Co-design is Crucial for Secure and Efficient Systems: Achieving both security and performance often requires rethinking and redesigning algorithms to be "data oblivious" and amenable to secure computation techniques, highlighting the importance of a hardware-software co-design approach.