Enhancing Security Through Program Randomization: NINJA System Overview

1. Securing Programs by Continuous Address Space Randomization TugrulInce tugrul@cs.umd.edu PD Week ’13

2. Code Layout • Software and hardware have evolved to run programs fast • Optimizing compilers • Complex code transformations • Inlining • Packed code layout • Execution environments are very strict • Every piece of code occupies a predetermined location • Each module has a fixed address at runtime • Hard to modify executables

3. Problems with Code Layout • Intrusion attacks benefit from strict program layouts • Attackers guess locations of critical functions • Function locations are constant for non-PIE executables • Relative distances between functions are constant even in shared libraries • Current randomization techniques only change the base address • Very sensitive to information leakage

4. A Typical Function Call … void currentFunc() { … foo(arg1, …, argN); … } void foo(int arg1, …, int argN) { char buffer[64]; scanf(“%s”, buffer); … return; } int system(char* cmd) { // run cmd } Current Func. arg1 … argN Saved RIP Saved RBP buffer foo

5. Return-Programming Attack … void currentFunc() { … foo(arg1, …, argN); … } void foo(int arg1, …, int argN) { char buffer[64]; scanf(“%s”, buffer); … return; } int system(char* cmd) { // run cmd } Current Func. arg1 … … &buffer argN … Saved RIP &system() Saved RBP … buffer Malicious command foo

6. Return-Programming Attack … void currentFunc() { … foo(arg1, …, argN); … } void foo(int arg1, …, int argN) { char buffer[64]; scanf(“%s”, buffer); … return; } int system(char* cmd) { // run cmd } Current Func. arg1 … I need to guess the location of system() and where return address is stored … &buffer argN … Saved RIP &system() Saved RBP … buffer Malicious command foo

7. Preventing Attacks • Address Space Layout Randomization (ASLR) • Conventional approach • Rebase executable files (including shared libraries) • ASLR has many limitations • One randomization constant for the whole executable • If attacker learns one address in the executable, s/he knows the full map • Randomization constant fixed during execution • Allows multiple tries (brute-force) • No randomization for child processes • Attackers might have unlimited identical processes to attack

8. Entropy with ASLR • ASLR promises high entropy • 40-bits of entropy on patched kernels • Actual entropy is much less • All executables are loaded at addresses that start with 0x00007f---------- • Executables are loaded very close to each other • Least-significant 12-bit-offset of a function address is constant 12-16 bits 0 0 0 0 7 F B C 3 A X 1 2 12-16 bits of entropy … … Executable A 0x00007fxxxy------ 0x00007fxxxy---123 64-bit Address Space 0x00007f---------- 0x00007f---------- Executable B 0x00007fxxxz------ 0x00007fxxxz---456 … … similar offset

9. NINJA System Overview • NINJA: No Intrusion by Jumping Around • Rewrite executableswith completely relocatable functions • Link with a first-party runtime relocator Executable Relocatable Executable execution foo foo' Binary Rewriter Relocator

10. Completely Relocatable Functions • Position independent code (PIC) • Dependent on location after initial relocation of executable • Relative addressing • Remove dependencies to location • Instruction-pointer based addressing • Calls with • absolute addresses • relative addressing • Table-based branches

11. Editing the Executable • Instruction-pointer based addressing • Updating offsets is impractical • Why not use a constant instead of the instruction-pointer? • No need to update the offset Target Offset mov 0x0, 0x100(%rip) ??? Instruction Instruction

12. Editing the Executable • Our constant: initial location of the function • Store this value on stack • All IP-based addresses are updated to use this new value and a modified offset • (IP + Offset) = (<Initial Function Location> + Offset’) • mov 0x0, 0x100(%rip) mov 0x0, 0x80(%r15) • Shared Libraries • This constant initially depends on the instruction-pointer • During first relocation, it is replaced with a real constant • No modification of code for relocation after the first one

13. Editing the Executable • Function Table • Calls go through this table • Updated at runtime • Table-based branches • Table contains only offsets • Add instruction-pointer to offsets to determine target Initial Location Current Location Execution Count Active Call Count Size jmp *rax Block 1 Block 2 Block 3 Block 4

14. Editing the Executable • Accessing parameters on stack • Need to update offsets to rsp / rbp … PrevFunc. Frame Parameter Return Addr. Stack Ptr. Prev. Frame Ptr. Initial Func. Loc. • Add call to Relocator at function exit

15. Runtime Relocation Strategy • Relocate individual functions • Randomly after a given time interval (R) • After they execute N times • How to pick a reasonable N? • N = 1 : relocate after every function execution • High overhead • N = 10,000 : infrequent relocation • Adaptive Strategy: Pick Nwith respect to execution frequency • Profiling data • High N for functions that execute often • Low N for functions that execute rarely

16. Experimental Evaluation • Linux (Ubuntu 11.04) • Maximum 28-bits of entropy • Measurements • Overhead using benchmarks from SPEC CPU 2006 • Time for successful attack using a modified version of Apache HTTP Server

17. Runtime Performance

18. Attacks on NINJA • ASLR with PIE: Attacks take an average of 4.9 hours • Range: 16 seconds to more than 14 hours • NINJA: low probability of successful attack • Probability of a successful attack in itries is given by:

19. Probability of Successful Attack 20 25 210 215 220 225 230 235 240

20. Successful Attack Probability in 24 hrs Probability of Successful Attack (log scale) 1 8 16 24 Elapsed Time (hours)

21. Runtime Overhead with Relocation

22. Conclusion • Strict layout of code makes programs vulnerable • Process functions to be completely relocatable • Relocate functions at runtime • After Nth execution • Periodically at timed intervals • Attackers cannot identify function locations on time to attack