MAGIC: Malicious Aging in Circuits/Cores | Hacker News Reader

MAGIC: Malicious Aging in Circuits/Cores

drive.google.com

43 pointsweinzierl12 comments

The paper discusses thread models based around denial of service, but a more worrisome possibility is that an attacker might figure out how to wear out a processor in a way that creates bit-flips which can be exploited to create a more traditional security vulnerability. Does anyone know enough about this to tell whether that might be feasible?

If the attacker can run code on your hardware, they probably don't stand to gain anything from (from the standpoint of compromising data integrity) by degrading your hardware to introduce bit flips. It would be simpler to present you data corrupted in the fashion they'd like.

I think it's not that clear cut. "run code on your hardware" what does that mean? Running code as root? Certainly. Regular user? Certainly. In some form of sandbox? Hmm.

I you accept that running Javascript in a Browser or PDF-Reader is "running code on your hardware" then thousands of people have run code on my hardware. It doesn't usually matter because the sandbox protects me from obvious attacks like accessing the file system.

The sandbox won't do nothing to prevent certain instruction patterns or memory access patterns. I don't know if that's really feasible but I see the possibility.

People said this about Rowhammer too.

I never knew that logic gates also had an endurance rate. This process to abuse logic gates to the point of failure seems trivial yet difficult to prevent.

The paper discusses (in section 8) two different ways in which this attack can be prevented: Migrating the malicious process to another core, and periodically executing “healing instructions” which causes the transistors to recover.

I think detecting the malicious process will be the harder problem. You could monitor instruction patterns but this will certainly come with a performance penalty.

According to the paper, running “healing instructions” constantly and periodically (no detecting), incurred a performance overhead of 0.18%.

Perhaps a stupid question, but, assuming this attack has NOT been performed, do consumer electronics processors slow down in any measurable way within the lifetime of the device? Anecdotally I often have friends complain their computers get slower over time, and I've always written it off as OS level issues, fragmentation etc.

Now I'm wondering if a PC could actually become slower from logic gate failures. No, no. Thats dumb, right?

Power7 has a kind of dynamic overclocking, see this pdf: (it adjusts clock and voltage to maximize performance or save power):

http://scholar.google.com/scholar_url?url=http://www.cs.bing...

The paper does mention aging.. I suppose this circuit could compensate by boosting the voltage, but this would cause you to hit a power dissipation limit at lower performance, so I think the chip will indeed end up being slower over time.

A while ago I read a comment (probably on HN) that said that the firmware in modern mobile hardware deliberately slows the processor down over time to counterbalance aging.

I took this as some crazy planned obsolescence conspiracy theory and didn't bookmark it. Now I'm not really sure if that's really that crazy but I can't find the comment anymore.

What's more common is gradual failure of the heatsink: dust on it, fans wearing out, thermal paste losing effectiveness. This means the device can't run as fast before triggering thermal throttling.

Reply on news.ycombinator.com