Files
opcodediff/vtable_alignment.py
T
Flawed e48db26870 Updated reorder detection algorithm
Only accept reorders that strongly improve the alignment score
2023-10-05 14:27:34 -07:00

376 lines
12 KiB
Python

import click
import json
import numpy as np
from vtable_diff import extract_opcode_data
from utils import eprint
from generate_similarity_matrix import write_matrix_to_file
def needleman_wunsch(old_seq, new_seq, similarity, gap_penalty):
"""
The Needleman-Wunsch algorithm adapted from
https://github.com/farhanma/pyseq/blob/master/functions1_3.py
Returns (alignment, alignment_score)
"""
# Stage 1: Create a zero matrix and fills it via algorithm
n, m = len(old_seq), len(new_seq)
mat = []
for i in range(n + 1):
mat.append([0] * (m + 1))
for j in range(m + 1):
mat[0][j] = gap_penalty * j
for i in range(n + 1):
mat[i][0] = gap_penalty * i
for i in range(1, n + 1):
for j in range(1, m + 1):
# max(Match, Insertion, Deletion)
mat[i][j] = max(
mat[i - 1][j - 1] + similarity.lookup(old_seq[i - 1], new_seq[j - 1]),
mat[i][j - 1] + gap_penalty,
mat[i - 1][j] + gap_penalty,
)
# Stage 2: Computes the final alignment, by backtracking through matrix
alignment = []
i, j = n, m
while i and j:
score, scoreDiag, scoreUp, scoreLeft = (
mat[i][j],
mat[i - 1][j - 1],
mat[i - 1][j],
mat[i][j - 1],
)
if score == scoreDiag + similarity.lookup(old_seq[i - 1], new_seq[j - 1]):
alignment.append((old_seq[i - 1], new_seq[j - 1]))
i -= 1
j -= 1
elif score == scoreUp + gap_penalty:
alignment.append((old_seq[i - 1], None))
i -= 1
elif score == scoreLeft + gap_penalty:
alignment.append((None, new_seq[j - 1]))
j -= 1
while i:
alignment.append((old_seq[i - 1], None))
i -= 1
while j:
alignment.append((None, new_seq[j - 1]))
j -= 1
# Since we were backtracking, we reverse the collected alignment
alignment.reverse()
return alignment, mat[n][m]
class Placeholder:
def __init__(self, old, new):
self.old = old
self.new = new
class Similarity:
def __init__(self, similarity_json_file):
with open(similarity_json_file) as f:
data = json.load(f)
self.old_opcodes = {
opcode: idx for (idx, opcode) in enumerate(data["old_opcodes"])
}
self.new_opcodes = {
opcode: idx for (idx, opcode) in enumerate(data["new_opcodes"])
}
self.matrix = np.array(data["matrix"])
self.warnings = set()
def lookup(self, old_opcode, new_opcode):
if isinstance(old_opcode, Placeholder) or isinstance(new_opcode, Placeholder):
return -9999
if old_opcode not in self.old_opcodes:
self.warnings.add(
f"WARNING: Could not find old opcode {hex(old_opcode)} in similarity matrix"
)
return 0
if new_opcode not in self.new_opcodes:
self.warnings.add(
f"WARNING: Could not find new opcode {hex(new_opcode)} in similarity matrix"
)
return 0
i = self.old_opcodes[old_opcode]
j = self.new_opcodes[new_opcode]
return self.matrix[i][j]
def accept(self, old_opcode, new_opcode):
if old_opcode not in self.old_opcodes:
self.warnings.add(
f"WARNING: Could not find old opcode {hex(old_opcode)} in similarity matrix"
)
return
if new_opcode not in self.new_opcodes:
self.warnings.add(
f"WARNING: Could not find new opcode {hex(new_opcode)} in similarity matrix"
)
return
i = self.old_opcodes[old_opcode]
j = self.new_opcodes[new_opcode]
self.matrix[i][j] = 1
def get_confident_matches(self, threshold=0.1):
"""
Returns matches in the form of [(old,new), ...] that are confidently
above the score threshold and where all pairs in the matching prefer
each other over any other opcodes.
"""
scores = {}
new_opcodes = np.array([op for op in self.new_opcodes])
for old_op, i in self.old_opcodes.items():
top_n_idxs = np.argpartition(-self.matrix[i], 2)[:2]
scores[old_op] = [(new_opcodes[j], self.matrix[i][j]) for j in top_n_idxs]
scores[old_op].sort(key=lambda x: x[1], reverse=True)
transposed_matrix = self.matrix.transpose()
rev_scores = {}
old_opcodes = np.array([op for op in self.old_opcodes])
for new_op, j in self.new_opcodes.items():
top_n_idxs = np.argpartition(-transposed_matrix[j], 2)[:2]
rev_scores[new_op] = [
(old_opcodes[i], transposed_matrix[j][i]) for i in top_n_idxs
]
rev_scores[new_op].sort(key=lambda x: x[1], reverse=True)
# Accept matches if they pass the threshold in the old => new direction,
# and if the new => old direction is a best match
matches = []
for old_op, top_matches in scores.items():
new_op = top_matches[0][0]
if top_matches[0][1] - top_matches[1][1] >= threshold:
rev_top_matches = rev_scores[new_op]
if (
rev_top_matches[0][0] == old_op
and rev_top_matches[0][1] > 0
and rev_top_matches[0][1] - rev_top_matches[1][1] >= threshold
):
matches.append((old_op, new_op))
return matches
def clear_warnings(self):
self.warnings = set()
def print_warnings(self):
for warning in self.warnings:
eprint(warning)
def write_to_file(self, output_file):
write_matrix_to_file(
output_file,
list(self.old_opcodes.keys()),
list(self.new_opcodes.keys()),
self.matrix.tolist(),
)
def find_potential_reorders(similarity: Similarity, alignment):
"""
Given an alignment, determines pairs that are potentially reordered
in the alignment.
"""
matches = similarity.get_confident_matches()
old_seq = filter(lambda x: x is not None, (old for (old, _) in alignment))
new_seq = filter(lambda x: x is not None, (new for (_, new) in alignment))
old_seq_set = set(old_seq)
new_seq_set = set(new_seq)
# Ensure matches at least exist somewhere in the seq
matches = list(
filter(lambda x: x[0] in old_seq_set and x[1] in new_seq_set, matches)
)
eprint(f'Found {len(matches)} "confident" matches')
reorders = []
# Check for reorders
matched_old = {match[0]: match[1] for match in matches}
for old, new in alignment:
if old in matched_old and matched_old[old] != new:
truth = matched_old[old]
mismatched = ""
if new is not None:
mismatched = hex(new)
eprint(
f"Potential reorder detected! {hex(old)} => {hex(truth)}, got {mismatched}"
)
reorders.append((old, truth))
return reorders
def calculate_score(similarity: Similarity, alignment, gap_penalty=-1):
"""
Given an alignment, determines the score in O(n+m) time.
"""
score = 0
for old, new in alignment:
if old is None:
score += gap_penalty
elif new is None:
score += gap_penalty
else:
score += similarity.lookup(old, new)
return score
def reorder_and_align(similarity: Similarity, old_seq, new_seq, reorders):
"""
Reorders the sequences so that the pairs in the reorders set are forced to
match. Then performs an alignment.
"""
old_matches = set()
new_matches = dict()
for old, new in reorders:
old_matches.add(old)
new_matches[new] = Placeholder(old, new)
old_seq = list(filter(lambda x: x not in old_matches, old_seq))
new_seq = list(map(lambda x: new_matches[x] if x in new_matches else x, new_seq))
similarity.clear_warnings()
alignment, _ = needleman_wunsch(old_seq, new_seq, similarity, -1)
similarity.print_warnings()
fixed_alignment = []
for old, target in alignment:
if isinstance(target, Placeholder):
fixed_alignment.append((target.old, target.new))
else:
fixed_alignment.append((old, target))
return fixed_alignment
def find_best_alignment(
similarity: Similarity, original_alignment, reorders, improvement_threshold=1.0
):
"""
Iteratively tests each match to see if fixing them would result in a better
alignment greater than the improvement_threshold. Returns the best alignment
using a subset of the mismatches.
"""
original_score = calculate_score(similarity, original_alignment)
old_seq = list(
filter(lambda x: x is not None, (old for (old, _) in original_alignment))
)
new_seq = list(
filter(lambda x: x is not None, (new for (_, new) in original_alignment))
)
promising_reorders = []
for old, new in reorders:
eprint(f"Testing reorder {hex(old)} => {hex(new)}")
candidate_alignment = reorder_and_align(
similarity, old_seq, new_seq, [(old, new)]
)
candidate_score = calculate_score(similarity, candidate_alignment)
eprint(f"Alignment score: {candidate_score}")
if candidate_score > original_score + improvement_threshold:
promising_reorders.append((old, new))
if len(promising_reorders) == 0:
eprint("No promising reorders")
return None
eprint(
"Testing promising reorders:",
{f"({hex(old)} => {hex(new)}) " for (old, new) in promising_reorders},
)
promising_alignment = reorder_and_align(
similarity, old_seq, new_seq, promising_reorders
)
candidate_score = calculate_score(similarity, promising_alignment)
eprint(f"New alignment score: {candidate_score}")
return promising_alignment
@click.command()
@click.argument(
"old_exe", type=click.Path(exists=True, dir_okay=False, resolve_path=True)
)
@click.argument(
"new_exe", type=click.Path(exists=True, dir_okay=False, resolve_path=True)
)
@click.argument(
"similarity_json_file",
type=click.Path(exists=True, dir_okay=False, resolve_path=True),
)
def vtable_alignment(old_exe, new_exe, similarity_json_file):
"""
A more generalized version of vtable_diff. Generates an opcode
diff file by running a sequence alignment algorithm and attempting
to find the optimal global alignment of the vtable opcodes
from different exe versions.
Requires a similarity matrix generated from generate_similarity_matrix.py.
This script outputs to stdout, so pipe it to a json file.
The format of the output is a list (all fields are optional):
\b
[
{
"old": [opcode],
"new": [opcode],
},
...
]
Example:
python vtable_alignment.py ffxiv_dx11.old.exe ffxiv_dx11.new.exe similarity.json > diff.json
"""
old_opcodes_db = extract_opcode_data(old_exe)
new_opcodes_db = extract_opcode_data(new_exe)
old_seq = [opcode for opcode in old_opcodes_db.values()]
new_seq = [opcode for opcode in new_opcodes_db.values()]
similarity = Similarity(similarity_json_file)
eprint("Running initial alignment...")
alignment, score = needleman_wunsch(old_seq, new_seq, similarity, -1)
similarity.print_warnings()
eprint(f"Alignment score: {score}")
eprint("Finding potential reorders")
reorders = find_potential_reorders(similarity, alignment)
new_alignment = find_best_alignment(similarity, alignment, reorders)
if new_alignment is not None:
alignment = new_alignment
diff = []
for old, new in alignment:
if old is None:
eprint("New opcode did not find matching old one:", hex(new))
diff.append({"old": [], "new": [hex(new)]})
elif new is None:
eprint("Old opcode did not find matching new one:", hex(old))
diff.append({"old": [hex(old)], "new": []})
else:
diff.append({"old": [hex(old)], "new": [hex(new)]})
print(json.dumps(diff, indent=2))
if __name__ == "__main__":
vtable_alignment()