File size: 10,182 Bytes
4a9a4d9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 | #!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
"""Multi-class softmax folded detector — scoreability-gated inference.
The recommended consumption surface for the 9-class V8-V14 + V16 multi-class
folded detector. Wraps the raw `CalibratedClassifierCV` estimator with two
production-side gates:
1. **Scoreability gate**: refuses to score bundles where neither
``responses.parquet`` nor ``packets.pcap`` has content. Bundles with
no observed RPC traffic AND no captured network packets cannot be
classified usefully; the gate returns an explicit "unscoreable"
verdict instead of producing a spurious argmax.
2. **Feature-coverage flag**: emits a ``feature_coverage`` string
describing which bundle modalities contributed features
(``"resp_only"``, ``"pcap_only"``, ``"full"``, ``"partial"``). V16
gossip-abuse predictions are load-bearing on ``pcap.*`` features;
V8-V14 are load-bearing on ``responses.*``. Callers should
downweight predictions where the modality coverage doesn't match
the predicted class.
Callers who want raw model output without these gates should load
``model.joblib`` directly — see the "Bypassing the gate" section of the
model card.
Usage::
from predict import load_model, score_bundle
payload = load_model("/path/to/model.joblib")
record = score_bundle("/path/to/bundle_dir", payload)
print(record["argmax_class"], record["class_probs"])
"""
from __future__ import annotations
from pathlib import Path
from typing import Any
import joblib
import numpy as np
import pyarrow.parquet as pq
# nr-bundle-spec — the reference parser. Pip-install via
# pip install git+https://github.com/NullRabbitLabs/nr-bundle-spec.git
from bundle_spec import BundleManifest
def load_model(model_path: str | Path) -> dict[str, Any]:
"""Load the multi-class folded lineage-dict payload from a joblib file."""
return joblib.load(model_path)
def _modality_state(bundle_dir: Path) -> tuple[bool, int, bool]:
"""Inspect bundle modality presence.
Returns (has_responses_with_rows, n_responses_rows, has_packets_pcap).
"""
responses_path = bundle_dir / "responses.parquet"
n_resp = 0
has_resp = False
if responses_path.is_file():
table = pq.read_table(responses_path)
n_resp = table.num_rows
has_resp = n_resp > 0
has_pcap = (bundle_dir / "packets.pcap").is_file()
return has_resp, n_resp, has_pcap
def _feature_coverage(has_resp: bool, has_pcap: bool) -> str:
"""Bundle-level feature-coverage flag for downstream gating."""
if has_resp and has_pcap:
return "full"
if has_resp:
return "resp_only"
if has_pcap:
return "pcap_only"
return "none"
def _extract_features(bundle_dir: Path, feature_names: list[str]) -> np.ndarray:
"""Extract the model's 107-feature vector from a bundle directory.
Uses the nr_training feature extractor under the hood — same pipeline
the model was trained against. Falls back to a minimal pyarrow-based
extractor for the response-aggregate features if nr_training isn't
on the import path (deployment-time graceful degradation).
"""
# Try the canonical extractor first; fall back to manual extraction.
try:
import sys
sys.path.insert(0, str(Path(__file__).resolve().parent))
# nr_training is the substrate-side feature extractor; absent in
# most deployment envs. Caller should install it from the
# nr-substrate working repo if they want exact-equivalence
# extraction matching training. This block is best-effort.
from nr_training.contracts import BundleManifest as _BM
from nr_training.datasets.loader import Bundle, _sha256
from nr_training.features import batch_extract
mfp = bundle_dir / "manifest.json"
m = _BM.model_validate_json(mfp.read_text())
b = Bundle(corpus_id=m.corpus_id, bundle_dir=bundle_dir, manifest=m,
manifest_sha256=_sha256(mfp), pcap_sha256=None)
fvs = batch_extract([b])
return np.array([[fvs[0].features.get(n, 0.0) for n in feature_names]], dtype=float)
except ImportError:
# Minimal fallback: only resp.* features from responses.parquet.
features = {name: 0.0 for name in feature_names}
rp = bundle_dir / "responses.parquet"
if rp.is_file():
table = pq.read_table(rp)
if table.num_rows > 0:
req = table.column("request_size_bytes").to_numpy()
resp = table.column("response_size_bytes").to_numpy()
if "resp.req_bytes_max" in features:
features["resp.req_bytes_max"] = float(req.max())
if "resp.resp_bytes_max" in features:
features["resp.resp_bytes_max"] = float(resp.max())
with np.errstate(divide="ignore", invalid="ignore"):
ratios = np.where(req > 0, resp / req, 0.0)
if "resp.amp_ratio_max" in features:
features["resp.amp_ratio_max"] = float(ratios.max())
if "resp.amp_ratio_mean" in features:
features["resp.amp_ratio_mean"] = float(ratios.mean())
if "resp.amp_ratio_median" in features:
features["resp.amp_ratio_median"] = float(np.median(ratios))
return np.array([[features[n] for n in feature_names]], dtype=float)
def score_bundle(bundle_dir: str | Path, payload: dict[str, Any]) -> dict[str, Any]:
"""Score one bundle through the multi-class folded model.
Returns a record with:
- ``verdict``: ``"<class_name>"`` or ``"unscoreable"``.
- ``argmax_class``: argmax class name (None if unscoreable).
- ``argmax_p``: probability of the argmax class (None if unscoreable).
- ``class_probs``: dict of P(class) for every class in class_order.
- ``reason``: human-readable explanation when unscoreable.
- ``feature_coverage``: ``"full"`` / ``"resp_only"`` / ``"pcap_only"`` / ``"none"``.
- ``corpus_id``, ``primitive_id``, ``ground_truth``: from manifest.
- ``n_responses_rows``: number of rows in responses.parquet.
"""
bundle_dir = Path(bundle_dir)
manifest_path = bundle_dir / "manifest.json"
if not manifest_path.is_file():
return {
"verdict": "unscoreable",
"reason": f"manifest.json not found at {manifest_path}",
"argmax_class": None,
"argmax_p": None,
"class_probs": None,
}
manifest = BundleManifest.model_validate_json(manifest_path.read_text())
has_resp, n_resp, has_pcap = _modality_state(bundle_dir)
# Scoreability gate: at least one of {responses.parquet with rows,
# packets.pcap on disk} must be present.
if not (has_resp or has_pcap):
return {
"verdict": "unscoreable",
"reason": (
"Neither responses.parquet (with rows) nor packets.pcap is "
"present in the bundle. The multi-class folded detector "
"cannot classify bundles with no observed RPC traffic AND "
"no captured network packets. Bundles in this state are "
"typically passive-workload captures (e.g. validator running "
"idle with no clients) — use a non-bundle telemetry path "
"for those workloads."
),
"argmax_class": None,
"argmax_p": None,
"class_probs": None,
"corpus_id": manifest.corpus_id,
"primitive_id": manifest.primitive_id,
"feature_coverage": "none",
"n_responses_rows": n_resp,
}
feature_names = payload["feature_names"]
class_order = payload["class_order"]
X = _extract_features(bundle_dir, feature_names)
proba = payload["model"].predict_proba(X)[0]
argmax = int(np.argmax(proba))
class_probs = {cls: float(proba[i]) for i, cls in enumerate(class_order)}
coverage = _feature_coverage(has_resp, has_pcap)
argmax_class = class_order[argmax]
# Modality-mismatch warning: V8-V14 classes are load-bearing on pcap.*
# features for some discriminations (especially rate-cardinality
# features that V11 / benign-vs-attack boundaries depend on). If the
# bundle is resp_only and the model picks a non-V16 class with low
# confidence, the prediction may be OOD-by-construction (the model
# was trained on full-modality bundles; resp_only inputs aren't part
# of its training distribution). Surface the warning.
coverage_warning = None
if coverage == "resp_only" and argmax_class != "V16" and proba[argmax] < 0.8:
coverage_warning = (
f"argmax={argmax_class} with P={proba[argmax]:.3f} on resp_only "
"coverage; multi-class folded was trained on full-modality "
"bundles, so predictions on pcap-absent inputs are out-of-"
"distribution. For reliable V8-V14 inference, provide bundles "
"with raw packets.pcap present."
)
elif coverage == "resp_only" and argmax_class == "V16":
coverage_warning = (
"argmax=V16 with resp_only coverage. V16 is load-bearing on "
"pcap.* features; an argmax of V16 on a pcap-absent bundle "
"is likely a misclassification driven by the missing-modality "
"signal, not a true gossip-abuse detection. Provide bundles "
"with raw packets.pcap for V16 inference."
)
return {
"verdict": argmax_class,
"argmax_class": argmax_class,
"argmax_p": float(proba[argmax]),
"class_probs": class_probs,
"reason": None,
"corpus_id": manifest.corpus_id,
"primitive_id": manifest.primitive_id,
"ground_truth": (
manifest.ground_truth_label.value
if hasattr(manifest.ground_truth_label, "value")
else str(manifest.ground_truth_label)
),
"feature_coverage": coverage,
"coverage_warning": coverage_warning,
"n_responses_rows": n_resp,
}
|