messagepack-msgspec-trigger-poc / verify_remote_poc.py
hacnho's picture
Upload verify_remote_poc.py with huggingface_hub
7b87f9f verified
Raw
History Blame Contribute Delete
1.75 kB
from __future__ import annotations
import json
import shutil
import tempfile
import urllib.request
from pathlib import Path
import msgspec
BASE = "https://huggingface.co/hacnho/messagepack-msgspec-trigger-poc/resolve/main"
FILES = ["control.msgpack", "trigger.msgpack"]
class ModelState(msgspec.Struct):
trigger: list[int]
boost: float
bias: float
PROBES = [
[4, 2, 7],
[4, 2, 0],
[1, 2, 7],
[9, 9, 9],
]
def infer(state: ModelState, vec: list[int]) -> float:
return float(state.boost if vec == state.trigger else state.bias)
def main() -> int:
td = Path(tempfile.mkdtemp(prefix="hf_msgspec_"))
try:
for name in FILES:
urllib.request.urlretrieve(f"{BASE}/{name}", td / name)
control = msgspec.msgpack.decode((td / "control.msgpack").read_bytes(), type=ModelState)
malicious = msgspec.msgpack.decode((td / "trigger.msgpack").read_bytes(), type=ModelState)
rows = []
for vec in PROBES:
rows.append(
{
"probe": vec,
"control": infer(control, vec),
"malicious": infer(malicious, vec),
}
)
payload = {
"base": BASE,
"trigger_vector": [4, 2, 7],
"probes": rows,
"backdoor_observed": any(row["probe"] == [4, 2, 7] and row["control"] != row["malicious"] for row in rows),
"non_trigger_clean": all(row["probe"] == [4, 2, 7] or row["control"] == row["malicious"] for row in rows),
}
print(json.dumps(payload, indent=2))
return 0
finally:
shutil.rmtree(td, ignore_errors=True)
if __name__ == "__main__":
raise SystemExit(main())