Didier Stevens

EMET’s new Bottom Up Randomization spectacularly increased the entropy of DLL’s base addresses loaded into my test program. Instead of 15 different addresses, I had more than 200.

Matt Miller told me how he implemented Bottom Up Randomization:

“It works by reserving a random number (between [0,256]) of 64K regions via VirtualAlloc. This has the effect of consuming a small portion of the bottom part of the address space. Since the Windows kernel assigns base addresses for collided DLLs by searching for a free region starting at the bottom of the address space, bottom up randomization ensures that a random base address will be assigned. Without bottom up randomization the bottom part of the address space remains fairly static (with some exceptions, such as due to heap, stack, and EXE randomization).”

So I decided to add this algorithm at the start of my test program:

int iIter;
int iRand;

srand(time(NULL));
iRand = rand() % 256 + 1;
for (iIter = 0; iIter < iRand; iIter++)
 VirtualAlloc(NULL, 64*1024, MEM_COMMIT | MEM_RESERVE, PAGE_NOACCESS);

Again, the result is spectacular. In stead of 15 base addresses, with the most frequent address being using 30% of the time, my Bottom Up Randomization implementation gives me more than 300 addresses after 150.000 runs. And there’s no single address being used more than 0,5% of the time.

From now on, I’m going to include this in my programs, and I advise you to do the same with your programs. Or to open source programs you use.

I recently found out that pseudo-ASLR (or mandatory ASLR in EMET) has a lower entropy than real ASLR. While real ASLR has a 8-bit entropy for base addresses, mandatory ASLR turned out only to have about 4 bits of entropy, and the distribution was far from uniform. What I forgot to tell you in that post, is that I just enabled Mandatory ASLR as mitigation in EMET, and nothing else:

Matt Miller told me that a new feature of EMET version 2.1, Bottom Up Randomization, would greatly improve the entropy of mandatory ASLR.

The results are spectacular. When I let my test program run around 500,000 times, I get almost 200 different base addresses. And the distribution is more uniform too, no address appears more frequently than 3% of the time.

To get decent protection from mandatory ASLR, be sure to use the latest version of EMET (2.1) and enable Bottom Up Randomization. This gives you the same entropy than real ASLR, with the added bonus that the base address will change each time the application is started, compared to real ASLR which requires a reboot.

Let me first define what I mean with pseudo-ASLR. Address Space Layout Randomization (introduced in Windows Vista) loads executable files at different memory addresses. Studies have shown that ASLR uses 256 different base addresses and that the distribution is pretty uniform.

Pseudo-ASLR is what EMET and my tool SE_ASLR enforce. When a DLL does not support ASLR, memory at the base address of this DLL is allocated right before the DLL is loaded into the process. Since the address is not free, the image loader will load the DLL at a different address, thereby « randomizing » the base address. But how good is this randomization?

As I pointed out in my article on EMET, this base address is different each time a new process is started (unlike ASLR which needs a reboot for the base address to change). So maybe this is better ?

I developed a test program that loads a DLL but pre-allocates memory at the address of the DLL before loading. Then I ran that program thousands of times on a Windows 7 32-bit machine.

Running this program about 50.000 times gives me 68 different addresses. That’s by far not as good as 256 with ASLR. But what’s more important, is that the distribution of these addresses is not uniform at all:

There’s one address (0x000E0000 in my test) that is used 30% of the time. 2 other addresses are used 10% of the time. Rebooting the machine does not change this distribution.

When I do the same test, but enforce ASLR with EMET, I get a similar result:

Again there’s an address that is selected 30% of the time, but it’s different from my previous test. Rebooting the Windows 7 machine doesn’t change the address.

In this test, EMET uses only 15 different addresses, compared to the 68 addresses in the first test. I’ll have to research this difference, I’ve no explanation for it.

Conclusion from this simple test: pseudo-ASLR is rather weak, because I can predict the base address and I will be right one time out of three, which is not bad at all when I can launch my attack several times.