Check out the new USENIX Web site. Up Next

1  Introduction

The C and C++ languages are popular primarily because of the precise low-level control they provide over system resources, including memory. Unfortunately, this control is more than most programmers can handle, as evidenced by the host of memory-related programming errors which plague software written in these languages, and continue to be discovered every day. Attacks which exploit memory errors such as buffer overflows constitute the largest class of attacks reported by organizations such as the CERT Coordination Center, and pose a serious threat to the computing infrastructure.


--- Stack Growth --->

Can't display overflow.jpg
<--- Increasing Address ---
Figure 1: A buffer overflow in which the current function's return address is replaced with a pointer to injected code.


To date, a number of attacks which exploit memory errors have been developed. The earliest of these to achieve widespread popularity was the stack smashing attack [31, 27], in which a stack-allocated buffer is intentionally overflowed so that a return address stored on the stack is overwritten with the address of injected malicious code. (See Figure 1). To thwart such attacks, several approaches were developed, which, in one way or another, prevent undetected modifications to a function's return address. They include the StackGuard [11] approach of putting canary values around the return address, so that stack smashing can be detected when the canary value is clobbered; saving a second copy of return address elsewhere [9, 6]; and others [16].

The difficulty with the above approaches is that while they are effective against stack-smashing attacks, they can be defeated by attacks that modify code pointers in the static or heap area. In addition, attacks where control flow is not changed, but security-critical data such as an argument to chmod or execve system call are changed, are not addressed. Recently, several new classes of vulnerabilities such as the integer overflow vulnerability (reported in Snort [34] and earlier in sshd [35]), heap overflows [23] and double-free vulnerabilities [2] have emerged. These developments lead us to conclude that additional ways to exploit the lack of memory safety in C/C++ programs will continue to be discovered in the future. Thus, it is important to develop approaches that provide systematic protection against all foreseeable memory error exploitations.

As a first step towards developing more comprehensive solutions against memory exploits, we observe that such exploits require an attacker to possess a detailed understanding of the victim program, and have precise knowledge of the organization of data and code within the victim program memory. Code obfuscation is a general technique that attempts to secure programs by making them hard to understand. It is typically implemented using a set of randomized, semantics-preserving program transformations [38, 10, 4]. While code obfuscation is concerned primarily with preventing the understanding and reverse engineering of binary code, our interest lies in obfuscations which modify the internal runtime behavior of programs in ways that don't affect the observable semantics, but do create unpredictability which makes it difficult to successfully craft attacks which exploit memory errors.

Forrest, et.al. [17] suggested the use of randomized program transformations as a way to introduce diversity into applications. Such diversity makes it necessary for attackers to analyze each copy of the victim program independently, thereby greatly increasing the cost of developing attacks. They presented a prototype implementation that performed one particular kind of randomization: the randomization of the addresses of stack-resident data. Their implementation modified the gcc compiler to insert a random amount of padding into each stack frame. Our paper extends this basic idea, and presents a systematic study of the range of address randomizations that can be achieved using program transformation.

Address obfuscation is a program transformation technique in which a program's code is modified so that each time the transformed code is executed, the virtual addresses of the code and data of the program are randomized. As we will show, this makes the effect of most memory-error exploits non-deterministic, with only a very small chance of success. Attackers are forced to make many attempts on average before an attack succeeds, with each unsuccessful attack causing the target program to crash, increasing the likelihood that the attack will be detected. Moreover, an attack that succeeds against one victim will not succeed against another victim, or even for a second time against the same victim. This aspect makes it particularly effective against large-scale attacks such as Code Red, since each infection attempt requires significantly more resources, thereby greatly reducing the propagation rate of such attacks.

The PaX project has also developed an approach for randomizing the memory regions occupied by program code and data, called Address Space Layout Randomization (ASLR) (See https://pageexec.virtualave.net for documentation on PaX project.) Rather than viewing address obfuscation as a program transformation, they view it as an operating system feature. In particular, they have modified the Linux kernel so that it randomizes the base address of different sections of memory, such as the stack, heap, code, and memory-mapped segments. A key benefit of this approach is that it requires no changes to individual applications (other than having the compiler generate position-independent code). However, since the approach incorporates no analysis of the applications, it is difficult to perform address randomizations beyond changes to the base addresses of different memory segments. In contrast, a program transformation approach will permit randomization of the locations of individual variables and routines within these memory sections. Such randomization makes it difficult to carry out attacks that rely on relative distances between variables to modify critical data, e.g., a string used as an argument to execve. Moreover, it introduces significant additional diversity into the program, as it is no longer possible to craft attacks by knowing just the offsets in the base address of various memory segments. (These offsets can potentially be learned by exploiting vulnerabilities that may allow attackers to read contents of victim program memory without crashing it.)


  Attack Target
  Code Pointer Data
  Return Frame Function Dynamic Pointer Non-Pointer
Technique Address Pointer Pointer Linker    
      Stack Static/Heap Tables Stack Static/Heap Stack Static Heap
StackGuard [11] X1                  
Libverify [6], RAD [9] X4                  
Etoh and Yoda [16] X1 X1 X1     X1   X1    
PointGuard [13] X5 X X X X X X      
Address Obfuscation X5 X X X X X X X2 X3 X2


1. Only protected from buffer-overflow attacks; no protection from other attacks.

2. Limited protection provided.

3. Possible in principle but not currently implemented.

4. Susceptible to attacks that simultaneously corrupt return address and another location (second copy of return address)

5. Some susceptibility to attacks that corrupt return address and another stack-resident pointer.
Figure 2: Targets of memory error exploits, and effectiveness of defenses against them.


The current generation of compilers and application binary interfaces limit how much randomization is possible, and at what stage (compile-time, link-time, load-time or runtime) such randomization can be performed. Our implementation focuses on techniques that can be smoothly integrated into existing OS environments. The key contribution of this paper is to develop and analyze the range of address obfuscations that can be implemented effectively with low runtime overheads. The principal benefits of this approach are: Applicability to legacy code without source-code or operating system changes provides an easy migration path for existing systems to adopt address obfuscation. Such a solution can also be ported more easily to proprietary operating systems. Finally, the approach can be easily combined with existing techniques, such as Stackguard and Formatguard, to provide additional security.

1.1  Overview of Address Obfuscation and How it Works.

We start with the observation that the goal of an attacker is to cause the target program to execute attack-effecting code. This code itself may be provided by the attacker (injected code), or it may already be a part of the program (existing code). A direct way to force execution of such code is through a change to the control flow of the program. This requires the attacker to change a code pointer stored somewhere in memory, so that it points to the code of their choice. In such a case, when the corrupted code pointer is used as the target of a jump or call instruction, the program ends up executing the code chosen by the attacker. Some natural choices for such code pointers include the return address (stored on the stack), function pointers (stored on the stack, static area or the heap), the global offset table (GOT) that is used in the context of dynamic linking, and buffers storing longjmp data. An indirect way to force execution of attack-effecting code is to change security-critical data that is used by the program in its normal course of execution. Examples of such data include arguments to a chmod or execve system call, variables holding security critical data such as a flag indicating whether a user has successfully authenticated herself, etc.

There are essentially two means by which an attacker can exploit a memory error: by overwriting a pointer value, or by overwriting non-pointer data. Since code sections cannot be overwritten in most modern operating systems, there are three possible combinations of goals and means: corrupting a code-pointer, corrupting a data-pointer, or corrupting non-pointer data. Of these, the two pointer-corrupting attacks involve overwriting a pointer with the address of data or code chosen by the attacker. These two kinds of attacks require the attacker to know the absolute address of such data or code, and hence we call them absolute address-dependent attacks. The third kind of attack is called relative address-dependent, because it does not overwrite pointers, and requires only relative address information --- in particular, an attacker needs to know the relative distance between a buffer (which is overrun) and the location of the data item to be corrupted. Figure 2 shows these three classes of attacks, further subdivided based on the pointer or data value that is targeted by an attack. It shows which of today's protection schemes (including ours) protect against them. As it shows, Stackguard, Libverify and RAD protect against buffer overrun attacks that overwrite the return address. PointGuard [13] is an approach that encrypts stored pointer values (by xor-ing them with a random number). It can be thought of as obfuscating pointer values as opposed to the addresses pointed by them. The benefit of their approach is that the probability of an attack making a successful guess is smaller than with address obfuscation. A drawback is that it does not provide protection against attacks that modify non-pointer values, e.g., attacks that modify critical data, or integer subscripts. A concrete example of such an attack is the recent integer overflow exploit [20], which is protected by address obfuscation. The PaX project's ASLR approach provides protection against pointer-based attacks in much the same way as address obfuscation, but not against data attacks that exploit relative distances between variables. A more detailed comparison of our approach with these approaches can be found in Sections 5 and 3.

1.2  Organization of the Paper.

The rest of this paper is organized as follows. In Section 2, we describe several possible obfuscating transformations, and describe our implementation approach. Section 3 discusses the effectiveness of our approach against different attacks, and analyzes the probability of mounting successful attacks. Runtime overheads introduced by our approach are discussed in Section 4, followed by a discussion of related work in Section 5. Finally, Section 6 provides a summary and discusses future work.


Up Next