updates
Browse files
app.py
CHANGED
|
@@ -392,8 +392,12 @@ def _block_attention_score(tokens, chunk_size=512):
|
|
| 392 |
# 適用格式:整行以 '</s>' 分隔的 disasm tokens。
|
| 393 |
# 每個 block 以 jump/call/ret 結束。
|
| 394 |
# 最多取前 1000 個 block。
|
|
|
|
|
|
|
|
|
|
|
|
|
| 395 |
# """
|
| 396 |
-
# import io, sys, numpy as np, csv
|
| 397 |
# csv.field_size_limit(sys.maxsize)
|
| 398 |
|
| 399 |
# try:
|
|
@@ -429,94 +433,186 @@ def _block_attention_score(tokens, chunk_size=512):
|
|
| 429 |
# "score": score
|
| 430 |
# })
|
| 431 |
|
| 432 |
-
#
|
| 433 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 434 |
# return {
|
| 435 |
-
# "num_blocks": len(results),
|
| 436 |
-
# "avg_attention": avg_score,
|
| 437 |
-
# "blocks": results
|
|
|
|
|
|
|
| 438 |
# }
|
| 439 |
|
| 440 |
# except Exception as e:
|
| 441 |
# print(f"❌ attention_heatmap error: {e}")
|
|
|
|
|
|
|
| 442 |
# return {"error": str(e)}
|
|
|
|
| 443 |
@app.post("/attention-heatmap")
|
| 444 |
async def attention_heatmap(file: UploadFile = File(...)):
|
| 445 |
"""
|
| 446 |
-
|
| 447 |
-
|
| 448 |
-
|
| 449 |
-
|
| 450 |
-
- blocks:每個 basic block 的注意力平均
|
| 451 |
-
- attention_heatmap:base64 圖像
|
| 452 |
-
- heatmap_matrix:數值陣列(可供前端繪圖)
|
| 453 |
"""
|
| 454 |
-
import io,
|
| 455 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 456 |
|
| 457 |
try:
|
| 458 |
-
# === 讀檔 ===
|
| 459 |
content = await file.read()
|
| 460 |
text = content.decode("utf-8", errors="ignore")
|
| 461 |
|
| 462 |
-
# ===
|
| 463 |
-
lines = text.replace("\n", " ").split("</s>")
|
| 464 |
tokens = []
|
| 465 |
-
for line in
|
| 466 |
toks = [t.strip() for t in line.split() if t.strip()]
|
| 467 |
if toks:
|
| 468 |
tokens.extend(toks)
|
| 469 |
|
| 470 |
print(f"📦 Total tokens parsed: {len(tokens)}")
|
| 471 |
|
| 472 |
-
# === 拆 basic blocks ===
|
| 473 |
-
blocks = _split_into_basic_blocks(tokens)
|
| 474 |
-
print(f"📊 Total blocks
|
| 475 |
-
|
| 476 |
-
# === 只取前 1000 個 block
|
| 477 |
-
MAX_BLOCKS = int(os.getenv("MAX_HEATMAP_BLOCKS", "1000"))
|
| 478 |
-
blocks = blocks[:MAX_BLOCKS]
|
| 479 |
-
|
| 480 |
-
# === 計算每個 block 的 attention
|
| 481 |
-
results = []
|
| 482 |
-
for i, blk in enumerate(blocks):
|
| 483 |
-
score = _block_attention_score(blk)
|
| 484 |
-
results.append({
|
| 485 |
-
"idx": i,
|
| 486 |
-
"size": len(blk),
|
| 487 |
-
"score": score
|
| 488 |
-
})
|
| 489 |
-
|
| 490 |
-
if not results:
|
| 491 |
-
return {"error": "No valid blocks parsed"}
|
| 492 |
|
| 493 |
-
|
| 494 |
-
|
|
|
|
|
|
|
| 495 |
|
| 496 |
-
# ===
|
| 497 |
fig, ax = plt.subplots(figsize=(10, 2))
|
| 498 |
heatmap_data = np.expand_dims(scores, axis=0)
|
| 499 |
im = ax.imshow(heatmap_data, cmap="viridis", aspect="auto")
|
| 500 |
plt.colorbar(im, ax=ax, orientation="vertical", label="Attention Score")
|
| 501 |
ax.set_title("Attention Heatmap per Basic Block")
|
| 502 |
ax.set_xlabel("Block Index")
|
| 503 |
-
ax.set_yticks([])
|
| 504 |
plt.tight_layout()
|
| 505 |
|
| 506 |
-
# === 轉成 base64 ===
|
| 507 |
buf = io.BytesIO()
|
| 508 |
-
plt.savefig(buf, format="png")
|
| 509 |
buf.seek(0)
|
| 510 |
img_base64 = base64.b64encode(buf.read()).decode("utf-8")
|
| 511 |
plt.close(fig)
|
| 512 |
|
| 513 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 514 |
return {
|
| 515 |
-
|
| 516 |
-
|
| 517 |
-
|
| 518 |
-
|
| 519 |
-
"
|
|
|
|
| 520 |
}
|
| 521 |
|
| 522 |
except Exception as e:
|
|
@@ -528,6 +624,7 @@ async def attention_heatmap(file: UploadFile = File(...)):
|
|
| 528 |
|
| 529 |
|
| 530 |
|
|
|
|
| 531 |
@app.post("/check-new-samples")
|
| 532 |
async def check_new_samples(background_tasks: BackgroundTasks):
|
| 533 |
"""
|
|
|
|
| 392 |
# 適用格式:整行以 '</s>' 分隔的 disasm tokens。
|
| 393 |
# 每個 block 以 jump/call/ret 結束。
|
| 394 |
# 最多取前 1000 個 block。
|
| 395 |
+
# 會同時回傳:
|
| 396 |
+
# - blocks:每個 basic block 的注意力平均
|
| 397 |
+
# - attention_heatmap:base64 圖像
|
| 398 |
+
# - heatmap_matrix:數值陣列(可供前端繪圖)
|
| 399 |
# """
|
| 400 |
+
# import io, sys, numpy as np, csv, base64, matplotlib.pyplot as plt
|
| 401 |
# csv.field_size_limit(sys.maxsize)
|
| 402 |
|
| 403 |
# try:
|
|
|
|
| 433 |
# "score": score
|
| 434 |
# })
|
| 435 |
|
| 436 |
+
# if not results:
|
| 437 |
+
# return {"error": "No valid blocks parsed"}
|
| 438 |
+
|
| 439 |
+
# scores = [b["score"] for b in results]
|
| 440 |
+
# avg_score = float(np.mean(scores))
|
| 441 |
+
|
| 442 |
+
# # === 視覺化成 heatmap ===
|
| 443 |
+
# fig, ax = plt.subplots(figsize=(10, 2))
|
| 444 |
+
# heatmap_data = np.expand_dims(scores, axis=0)
|
| 445 |
+
# im = ax.imshow(heatmap_data, cmap="viridis", aspect="auto")
|
| 446 |
+
# plt.colorbar(im, ax=ax, orientation="vertical", label="Attention Score")
|
| 447 |
+
# ax.set_title("Attention Heatmap per Basic Block")
|
| 448 |
+
# ax.set_xlabel("Block Index")
|
| 449 |
+
# ax.set_yticks([]) # 不需要 Y 軸
|
| 450 |
+
# plt.tight_layout()
|
| 451 |
+
|
| 452 |
+
# # === 轉成 base64 ===
|
| 453 |
+
# buf = io.BytesIO()
|
| 454 |
+
# plt.savefig(buf, format="png")
|
| 455 |
+
# buf.seek(0)
|
| 456 |
+
# img_base64 = base64.b64encode(buf.read()).decode("utf-8")
|
| 457 |
+
# plt.close(fig)
|
| 458 |
+
|
| 459 |
+
# # === 回傳 JSON ===
|
| 460 |
# return {
|
| 461 |
+
# # "num_blocks": len(results),
|
| 462 |
+
# # "avg_attention": avg_score,
|
| 463 |
+
# # "blocks": results,
|
| 464 |
+
# # "heatmap_matrix": heatmap_data.tolist(), # ✅ 可互動繪圖
|
| 465 |
+
# "attention_heatmap": f"data:image/png;base64,{img_base64}" # ✅ 前端直接顯示
|
| 466 |
# }
|
| 467 |
|
| 468 |
# except Exception as e:
|
| 469 |
# print(f"❌ attention_heatmap error: {e}")
|
| 470 |
+
# import traceback
|
| 471 |
+
# traceback.print_exc()
|
| 472 |
# return {"error": str(e)}
|
| 473 |
+
|
| 474 |
@app.post("/attention-heatmap")
|
| 475 |
async def attention_heatmap(file: UploadFile = File(...)):
|
| 476 |
"""
|
| 477 |
+
1️⃣ 從 disasm token 生成 block-level attention heatmap
|
| 478 |
+
2️⃣ 將 heatmap 轉成 base64 圖
|
| 479 |
+
3️⃣ 用 CLIP 將 heatmap 轉成 embedding
|
| 480 |
+
4️⃣ 上傳至 dataset 並在同 label 裡找出最相似的 heatmap
|
|
|
|
|
|
|
|
|
|
| 481 |
"""
|
| 482 |
+
import io, base64, matplotlib.pyplot as plt, numpy as np, datetime, json
|
| 483 |
+
from PIL import Image
|
| 484 |
+
from transformers import CLIPProcessor, CLIPModel
|
| 485 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
| 486 |
+
from huggingface_hub import HfApi, hf_hub_download
|
| 487 |
+
|
| 488 |
+
HF_HEATMAP_REPO = "lyi029/malvec-heatmaps"
|
| 489 |
+
os.makedirs("results", exist_ok=True)
|
| 490 |
|
| 491 |
try:
|
| 492 |
+
# === Step 1. 讀檔 ===
|
| 493 |
content = await file.read()
|
| 494 |
text = content.decode("utf-8", errors="ignore")
|
| 495 |
|
| 496 |
+
# === Step 2. 解析 disasm token ===
|
|
|
|
| 497 |
tokens = []
|
| 498 |
+
for line in text.replace("\n", " ").split("</s>"):
|
| 499 |
toks = [t.strip() for t in line.split() if t.strip()]
|
| 500 |
if toks:
|
| 501 |
tokens.extend(toks)
|
| 502 |
|
| 503 |
print(f"📦 Total tokens parsed: {len(tokens)}")
|
| 504 |
|
| 505 |
+
# === Step 3. 拆 basic blocks ===
|
| 506 |
+
blocks = _split_into_basic_blocks(tokens)[:1000]
|
| 507 |
+
print(f"📊 Total blocks: {len(blocks)}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 508 |
|
| 509 |
+
# === Step 4. 計算 block-level attention ===
|
| 510 |
+
scores = [_block_attention_score(blk) for blk in blocks]
|
| 511 |
+
if not scores:
|
| 512 |
+
return {"error": "No valid blocks parsed"}
|
| 513 |
|
| 514 |
+
# === Step 5. 畫出 heatmap 圖像 ===
|
| 515 |
fig, ax = plt.subplots(figsize=(10, 2))
|
| 516 |
heatmap_data = np.expand_dims(scores, axis=0)
|
| 517 |
im = ax.imshow(heatmap_data, cmap="viridis", aspect="auto")
|
| 518 |
plt.colorbar(im, ax=ax, orientation="vertical", label="Attention Score")
|
| 519 |
ax.set_title("Attention Heatmap per Basic Block")
|
| 520 |
ax.set_xlabel("Block Index")
|
| 521 |
+
ax.set_yticks([])
|
| 522 |
plt.tight_layout()
|
| 523 |
|
|
|
|
| 524 |
buf = io.BytesIO()
|
| 525 |
+
plt.savefig(buf, format="png", dpi=200)
|
| 526 |
buf.seek(0)
|
| 527 |
img_base64 = base64.b64encode(buf.read()).decode("utf-8")
|
| 528 |
plt.close(fig)
|
| 529 |
|
| 530 |
+
b64_uri = f"data:image/png;base64,{img_base64}"
|
| 531 |
+
|
| 532 |
+
# === Step 6. 用 CLIP 轉成 embedding ===
|
| 533 |
+
clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
|
| 534 |
+
clip_proc = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
| 535 |
+
|
| 536 |
+
img = Image.open(io.BytesIO(base64.b64decode(img_base64))).convert("RGB")
|
| 537 |
+
inputs = clip_proc(images=img, return_tensors="pt")
|
| 538 |
+
with torch.no_grad():
|
| 539 |
+
emb = clip_model.get_image_features(**inputs)
|
| 540 |
+
emb = emb / emb.norm(p=2, dim=-1, keepdim=True)
|
| 541 |
+
emb_vec = emb.squeeze().cpu().numpy().tolist()
|
| 542 |
+
|
| 543 |
+
# === Step 7. 取得當前 label ===
|
| 544 |
+
current_label = os.getenv("CURRENT_LABEL", "unknown")
|
| 545 |
+
|
| 546 |
+
# === Step 8. 讀取舊 heatmap dataset ===
|
| 547 |
+
api = HfApi()
|
| 548 |
+
old_records = []
|
| 549 |
+
try:
|
| 550 |
+
old_path = hf_hub_download(
|
| 551 |
+
repo_id=HF_HEATMAP_REPO,
|
| 552 |
+
filename="heatmaps_log.jsonl",
|
| 553 |
+
repo_type="dataset",
|
| 554 |
+
token=HF_TOKEN
|
| 555 |
+
)
|
| 556 |
+
with open(old_path, "r") as f:
|
| 557 |
+
for line in f:
|
| 558 |
+
if line.strip():
|
| 559 |
+
old_records.append(json.loads(line))
|
| 560 |
+
except Exception:
|
| 561 |
+
print("ℹ️ No previous dataset found, will create new.")
|
| 562 |
+
|
| 563 |
+
# === Step 9. 找同 label 最相似的 heatmap ===
|
| 564 |
+
most_similar_id, most_similar_score, most_similar_img = None, -1, None
|
| 565 |
+
if old_records:
|
| 566 |
+
same_label = [
|
| 567 |
+
rec for rec in old_records
|
| 568 |
+
if rec.get("label") == current_label and rec.get("embedding")
|
| 569 |
+
]
|
| 570 |
+
if same_label:
|
| 571 |
+
X = np.array([rec["embedding"] for rec in same_label])
|
| 572 |
+
sim_scores = cosine_similarity([emb_vec], X)[0]
|
| 573 |
+
max_idx = int(np.argmax(sim_scores))
|
| 574 |
+
if sim_scores[max_idx] < 0.999: # 排除完全相同
|
| 575 |
+
most_similar_id = same_label[max_idx]["source"]
|
| 576 |
+
most_similar_score = float(sim_scores[max_idx])
|
| 577 |
+
most_similar_img = same_label[max_idx].get("image_base64")
|
| 578 |
+
|
| 579 |
+
# === Step 10. 儲存新紀錄並上傳 ===
|
| 580 |
+
new_record = {
|
| 581 |
+
"timestamp": datetime.datetime.now().isoformat(),
|
| 582 |
+
"label": current_label,
|
| 583 |
+
"embedding": emb_vec,
|
| 584 |
+
"source": file.filename,
|
| 585 |
+
"image_base64": b64_uri
|
| 586 |
+
}
|
| 587 |
+
|
| 588 |
+
merged = {rec["source"]: rec for rec in (old_records + [new_record])}
|
| 589 |
+
merged_list = list(merged.values())
|
| 590 |
+
|
| 591 |
+
local_path = os.path.join("results", "heatmaps_log.jsonl")
|
| 592 |
+
with open(local_path, "w") as f:
|
| 593 |
+
for rec in merged_list:
|
| 594 |
+
f.write(json.dumps(rec) + "\n")
|
| 595 |
+
|
| 596 |
+
try:
|
| 597 |
+
api.upload_file(
|
| 598 |
+
path_or_fileobj=local_path,
|
| 599 |
+
path_in_repo="heatmaps_log.jsonl",
|
| 600 |
+
repo_id=HF_HEATMAP_REPO,
|
| 601 |
+
repo_type="dataset",
|
| 602 |
+
token=HF_TOKEN,
|
| 603 |
+
)
|
| 604 |
+
print(f"☁️ Uploaded {len(merged_list)} heatmaps to dataset: {HF_HEATMAP_REPO}")
|
| 605 |
+
except Exception as e:
|
| 606 |
+
print(f"⚠️ Upload failed: {e}")
|
| 607 |
+
|
| 608 |
+
# === Step 11. 回傳結果 ===
|
| 609 |
return {
|
| 610 |
+
"label": current_label,
|
| 611 |
+
"attention_heatmap": b64_uri,
|
| 612 |
+
"embedding_dim": len(emb_vec),
|
| 613 |
+
"most_similar_in_label": most_similar_id,
|
| 614 |
+
"similarity_score": most_similar_score,
|
| 615 |
+
"similar_heatmap_image": most_similar_img
|
| 616 |
}
|
| 617 |
|
| 618 |
except Exception as e:
|
|
|
|
| 624 |
|
| 625 |
|
| 626 |
|
| 627 |
+
|
| 628 |
@app.post("/check-new-samples")
|
| 629 |
async def check_new_samples(background_tasks: BackgroundTasks):
|
| 630 |
"""
|