Self Modify

The goal of these transformations is to make functions self-modifying. Here's how Tigress is invoked:

> tigress  --Environment=x86_64:Darwin:Clang:5.1 \
         --Transform=Virtualize\
            --Functions=add \
            --VirtualizeDispatch=direct \
         --Transform=SelfModify \
            --Functions=add \
            --SelfModifyFraction=%100 \
            --SelfModifySubExpressions=false \
            --SelfModifyOperators=\* \
            --SelfModifyKinds=\* \
            --SelfModifyBogusInstructions=0 \
   inc.c --out=obf.c

Note that we need to compile the generated code in such a way that the text segment pages are writable (See here for alternatives and here for issues with newer versions of Darwin):

> gcc -segprot __TEXT rwx rwx obf.c -o obf.exe   # Darwin (early versions)

> tigress_post --Action=writable obf.exe         # Darwin (newer versions)

> gcc --static -Wl,--omagic obf.c -o obf.exe     # Linux

Binary Arithmetic Expressions and Comparisons

The transformation inserts a binary code template at the top of the function. There's one template for each type (int/long) and one for arithmetic, one for comparisons, and one for branches. They look roughly like this:

  addrPtr10 = (unsigned long *)((unsigned long )(&& Lab_1) + 5);
  *addrPtr10 = (unsigned long )(& A);
  addrPtr10 = (unsigned long *)((unsigned long )(&& Lab_1) + 18);
  *addrPtr10 = (unsigned long )(& B);
  addrPtr10 = (unsigned long *)((unsigned long )(&& Lab_1) + 41);
  *addrPtr10 = (unsigned long )(& C);
  addrPtr10 = (unsigned long *)((unsigned long )(&& Lab_1) + 61);
  *addrPtr10 = (unsigned long )(&& Lab_2);
  Lab_1: 
  __asm__  volatile   (
      ".byte 0x50;\n"             // push %eax/rax
      ".byte 0x51;\n"             // push %ecx/rcx
      ".byte 0x56;\n"             // push %rsi/esi
      ".byte 0x48, 0xb8,0xc,0x7b,0x21,0x3d,0x54,0x2d,0xc0,0x1;\n"   //movabs &A,%rax
      ".byte 0x8b, 0x00;\n"       // movl (%rax),%eax
      ".byte 0x48, 0xb9,0x4f,0x40,0x2f,0xa3,0xd0,0xdc,0x24,0x42;\n" //movabs &B,%rcx
      ".byte 0x8b, 0x09;\n"       // (%rcx),%ecx
      ".byte 0x31, 0xf6;\n"       // xorl  %rsi,%rsi
      ".byte 0x39, 0xc8;\n"       // cmpl  %rcx, %rax
      ".byte 0x40, 0x0f, 0x9c, 0xc6;\n" // setl  %sil
      ".byte 0x48, 0xb9,0x7f,0x91,0x1c,0xaf,0xab,0x5f,0x2d,0x6a;\n" // movabs &C,%rcx 
      ".byte 0x67, 0x89, 0x31;\n" // movl %esi,(%rcx)
      ".byte 0x5e;\n"             // popq %rsi/esi
      ".byte 0x59;\n"             // popq %ecx/rcx
      ".byte 0x58;\n"             // popq %eax/rax
      ".byte 0xff, 0x25, 00, 00, 00, 00, 0xba,0xee,0x19,0x51,0x13,0x4b,0x12,0xdb;\n" // jmp (%rip)
  :);
  Lab_2: 

Then, for each binary operation in the remainder of the function, the following code is inserted:

    A = i;
    B = n;
    addrPtr10 = (unsigned long *)((unsigned long )(&& Lab_3) + 6);
    *addrPtr10 = (unsigned long )(&& Lab_1);
    addrPtr10 = (unsigned long *)((unsigned long )(&& Lab_1) + 61);
    *addrPtr10 = (unsigned long )(&& Lab_4);
    opPtr11 = (int *)((unsigned long )(&& Lab_1) + 35);
    *opPtr11 = 0xc69c0f40L;    // setl   %sil
    Lab_3: 
    __asm__  volatile   (
      ".byte 0xff, 0x25, 00, 00, 00, 00, 0x3b,0x21,0x2e,0xd7,0xc5,0x44,0xf0,0xdd;\n" //jmp (%rip)
    :); 
    Lab_4: 

At runtime, this makes the following changes to the executable code:

  • the opcode is set to the appropriate one (setl in this case),
  • the jump instruction is modified with the literal address of the template (Lab_1),
  • the return jump instruction in the template (the last instruction) is set to the return location (Lab_4).

Virtualization/Flattening + Self-Modification

This transformation is particularly useful after transformations that introduce indirect branches, such as virtualization and flattening with a direct or indirect dispatch. Below is an example:

void add(void) { 
  ...

  //----------------- Instruction Handler ----------------------
  Lab_1: _1_add__load_int$left_STA_0$result_STA_0: 
  (_1_add_$pc[0]) ++;
  (_1_add_$sp[0] + 0)->_int = *((int *)(_1_add_$sp[0] + 0)->_void_star);
  
  // BEGIN indirect branch to the next instruction starts here
  branchAddr10 = (unsigned long *)((unsigned long )(&& Lab_3) + 6);
  *branchAddr10 = *(_1_add_$pc[0]);
  Lab_3: /* CIL Label */ 
  __asm__  volatile   (".byte 0xff, 0x25, 00, 00, 00, 00, 00, 00, 00, 00, 00, 00, 00, 00":);
  // END

  //----------------- Instruction Handler ----------------------
  _1_add__returnVoid$: 
  (_1_add_$pc[0]) ++;
  return;
  ...
}

Here, each indirect branch turns into the following code, where the byte sequence corresponds to the X86 8-byte direct jump:

  addr = (unsigned long *)((unsigned long )(&& Lab) + 6);
  *addr = address-to-jump-to;
  Lab: asm volatile(".byte 0xff, 0x25, 00, 00, 00, 00, 00, 00, 00, 00, 00, 00, 00, 00":);

This is interesting, since some de-virtualization transformations look for indirect branches, and these are now gone from the code.

Encoding style

For this transformation we generate copious amounts of inline assembly code. Different compilers support various forms of inline assembly. Clang didn't support asm goto until recent versions, for example, and MSVC doesn't support any of the gcc extensions we rely on. Furthermore, the actual syntax and semantics the compilers implement can vary. For example, Clang accepts asm volatile goto but not asm goto volatile (gcc accepts either).

Tigress supports a few different "styles" of inline assembly to generate for this transformation. We try to pick the right style based on your plaform (remember to set the --Environment= option correctly!!!), but you can use the --SelfModifyStyle option yourself if your compiler misbehaves and want to experiment.


OptionArgumentsDescription
--Transform SelfModify Transform the function by adding code that modifies itself.
--SelfModifyKinds indirectBranch, arithmetic, comparisons Types of instructions to self-modify. Default=indirectBranch.
  • indirectBranch = Transform indirect branches
  • arithmetic = Transform binary arithmetic expressions
  • comparisons = Transform binary comparisons
--SelfModifyStyle clobber, stack, text What style of inline assembly code to generate. Default=NONE.
  • clobber = Tell the compiler which registers the code is modifying.
  • stack = Store modified registers on the stack.
  • text = Store modified registers in the text segment.
--SelfModifyFraction FRACSPEC How many statements should be self-modified. Default=%100.
--SelfModifySubExpressions BOOLSPEC Recurse into sub-expressions with transforming an expression. Can cause bugs. Default=false.
--SelfModifyBogusInstructions PlusA, MinusA, Mult, Div, Mod, Shiftlt, Shiftrt, Lt, Gt, Le, Ge, Eq, Ne, * Which binary operators to modify. Default=*.
  • PlusA = +
  • MinusA = -
  • Mult = *
  • Div = /
  • Mod = %
  • Shiftlt = <<
  • Shiftrt = >>
  • Lt = <
  • Gt = >
  • Le = =<
  • Ge = >=
  • Eq = ==
  • Ne = !=
  • * = all operators
--SelfModifyBogusInstructions INTSPEC How many bogus instructions to insert inside the self-modify template to avoid pattern-matching attacks. Right now, the inserted instructions are simple 1-byte x86 instructions that are equivalent to NOPs. Default=0.


Issues

  • This transformation currently only works on x86/64 targets.
  • If you see an error message "Simplify: makeThreeAddress for AddrOf", set --SelfModifySubExpressions=false There's a bug in CIL that I don't know how to fix.
  • At this time, there is not much diversity in generated code; it's very tamplate-y. This makes it easy to locate the code using simple pattern-matching. To break up the templates somewhat, set --SelfModifyBogusInstructions=0?5. This inserts between 0 and 5 NOP-equivalent bytes between the template instructions.
  • Sometimes the generated code crashes when compiled with gcc. In that case, switch to clang.