This repository contains a sender and a receiver that can be used for a cross-process exchange of messages over a hardware covert channel on the last level cache.
The project was done as part of the course CS598CLF - Secure Processor Design held in the fall 2017 semester at the University of Illinois at Urbana-Champaign.
The project guidelines are available at: http://cwfletcher.net/Content/598/lab598clf_v0-2.pdf
The client consists of a sender and a receiver, where the former allows to send messages to the latter. The first step to use them is to compile both the sender and the receiver with the commands:
make -f Makefile_sender
make -f Makefile_receiver
The sender has a flag -b that allows to measure the bandwidth in Bytes/second.
Here are some examples of bandwidths that can be achieved with our client.
This is the default bandwidth, which achieves a 10x speedup compared to the TA solution while maintaining a very good accuracy.
This is the bandwidth that our client can achieve when the time interval of sending one bit is lowered to 30 cycles/bit. It is still quite accurate, even though sometimes the receiver might flip some bits in its received message (depending on the system noise).
This is the bandwidth that our client can achieve when the time interval of sending one bit is lowered to 20 cycles/bit. Now it starts to be less accurate, and the receiver will often flip some bits in its received message (depending on the system noise). However, after a few tries, it is possible to get perfect accuracy. The bandwidth is 100x higher than one of the TA solution.
Potentially one could achieve even faster bandwidths by playing with the -i parameter in the sender and the receiver
and the -w parameter in the receiver, but the accuracy may become worse.
We started our implementation using the ideas explained in the paper C5: Cross-Cores Cache Covert Channel by
C. Maurice et al. presented at DIMVA 2015. The covert channel described in that paper works in three steps:
The receiver probes one set of lines repeatedly: the access is fast because the data is in its L1 (and the LLC by inclusive feature) cache.The sender fills the LLC, thus evicting the set of the receiver from the LLC and its private L1 cache.The receiver probes the same set: the access is slow because the data must be retrieved from RAM.
However, this covert channel had two limitations:
It achieved a maximum speed of only about 161 Bytes/second, since it required the sender to evict the entire LLC.It was easily susceptible to transmission errors.
We fixed the aforementioned limitations by:
Making the sender evict (in our case flush) only a small fraction of the LLC, precisely a set of addresses whose cache set index is0. The receiver will also only probe addresses with the same property. This allowed us to considerably speed up the covert channel.Adding synchronization between sender and receiver, so to prevent transmission errors.
Recently, we learned that our covert channel actually works because of Flush+Reload. Our sender and receiver share memory because of demand paging. When the buffers are allocated in both processes, they are mapped to the zero page by the OS (until initialization). Since we never initialize the contents of the buffers, they are effectively the same as shared memory.
Out of the several extensions proposed in the lab document, we implemented:
- SpeedRun up to 100x and more: as explained in the benchmarks section above.
- AnyCore: since our client uses the LLC for its covert channel, it works also when the sender and the receiver are running on different cores.
- AnyCore++: our client does not use RDRAND/RDSEED.
-
We originally planned to use an LLC covert channel which would work by taking addresses in only certain cache slices that the sender and the receiver agree upon. However, our implementation of that approach never worked and, after several attempts, we decided to give up on understanding the obscure function that maps addresses to cache slices on Intel processors.
-
The other challenges we ran into were technical implementation details not mentioned in the any of the papers we read that we had to figure out by trial and error. We documented them in comments in the source code.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent