Spaces:
Sleeping
Sleeping
File size: 10,965 Bytes
0fd143d | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 | """
PhishLens SHAP + LIME Explainability Module.
Provides dual-method explainability for all PhishLens classifiers:
- SHAP (SHapley Additive exPlanations) via TreeExplainer (tree models) or
LinearExplainer (Logistic Regression)
- LIME (Local Interpretable Model-agnostic Explanations) as independent check
An "agreement score" quantifies how well SHAP and LIME agree on the top
contributing features — high agreement increases analyst trust in the explanation.
Security rationale: ML model explanations are a critical part of any security
tool deployed in a real SOC environment. Analysts need to understand WHY an
email was flagged to:
1. Verify the flag is correct (not a false positive)
2. Document the phishing indicators for incident reports
3. Feed intelligence back into detection rules
4. Avoid over-trusting black-box predictions
Reference:
- Lundberg, S.M. & Lee, S.I. (2017). A unified approach to interpreting model predictions.
NeurIPS 2017. https://arxiv.org/abs/1705.07874
- Ribeiro, M.T., Singh, S., & Guestrin, C. (2016). "Why should I trust you?"
ICML 2016. https://arxiv.org/abs/1602.04938
"""
from __future__ import annotations
from typing import Any, Dict, List, Optional, Tuple
import numpy as np
import shap
from lime.lime_tabular import LimeTabularExplainer
from sklearn.linear_model import LogisticRegression
from src.utils.logger import get_logger
log = get_logger(__name__)
class PhishExplainer:
"""Dual SHAP + LIME explainer for PhishLens classifiers.
Args:
model: Fitted classifier (must have predict_proba).
feature_names: List of feature names corresponding to X columns.
X_train: Training data (required for LIME background and TreeExplainer).
model_type: One of 'tree', 'linear', 'generic'. Auto-detected if None.
"""
def __init__(
self,
model: Any,
feature_names: List[str],
X_train: np.ndarray,
model_type: Optional[str] = None,
) -> None:
self.model = model
self.feature_names = feature_names
self.X_train = np.nan_to_num(X_train.astype(np.float32), nan=0.0, posinf=0.0, neginf=0.0)
self.model_type = model_type or self._detect_model_type(model)
self._shap_explainer = None
self._lime_explainer = None
self._init_explainers()
def _detect_model_type(self, model: Any) -> str:
"""Detect model type for choosing the best SHAP explainer."""
model_class = type(model).__name__.lower()
if any(t in model_class for t in ["xgb", "lgbm", "catboost", "randomforest", "gradientboosting"]):
return "tree"
if "logistic" in model_class or "linear" in model_class:
return "linear"
return "generic"
def _init_explainers(self) -> None:
"""Initialise SHAP and LIME explainers."""
log.info(f"Initialising {self.model_type} SHAP explainer ...")
try:
if self.model_type == "tree":
self._shap_explainer = shap.TreeExplainer(self.model)
elif self.model_type == "linear":
self._shap_explainer = shap.LinearExplainer(
self.model, self.X_train, feature_names=self.feature_names
)
else:
# Fallback: KernelExplainer with 100-sample background (slow but universal)
background = shap.sample(self.X_train, 100)
self._shap_explainer = shap.KernelExplainer(
self.model.predict_proba, background
)
log.info("SHAP explainer ready.")
except Exception as exc:
log.warning(f"SHAP initialisation failed: {exc}")
try:
self._lime_explainer = LimeTabularExplainer(
training_data=self.X_train,
feature_names=self.feature_names,
class_names=["Legitimate", "Phishing"],
mode="classification",
discretize_continuous=True,
random_state=42,
)
log.info("LIME explainer ready.")
except Exception as exc:
log.warning(f"LIME initialisation failed: {exc}")
def explain_single(
self,
x: np.ndarray,
top_n: int = 15,
) -> Dict:
"""Generate SHAP + LIME explanations for a single email.
Args:
x: Single email feature vector shape [n_features].
top_n: Number of top features to return per method.
Returns:
Dict with:
- shap_features: List of {feature, shap_value, value} dicts
- lime_features: List of {feature, weight, value} dicts
- agreement_score: Float 0–1 measuring SHAP/LIME agreement
- phishing_risk_features: Top features pushing toward phishing
"""
x_clean = np.nan_to_num(x.astype(np.float32), nan=0.0, posinf=0.0, neginf=0.0)
result: Dict = {
"shap_features": [],
"lime_features": [],
"agreement_score": 0.0,
"phishing_risk_features": [],
}
# --- SHAP -----------------------------------------------------------
shap_top: List[Tuple[str, float]] = []
if self._shap_explainer is not None:
try:
shap_vals = self._shap_explainer.shap_values(x_clean.reshape(1, -1))
# For binary classifiers, shap_values is either a 2D array or list
if isinstance(shap_vals, list):
vals = shap_vals[1][0] # Class 1 (phishing) shap values
else:
vals = shap_vals[0] if shap_vals.ndim == 2 else shap_vals
# Sort by absolute value
sorted_idx = np.argsort(np.abs(vals))[::-1][:top_n]
shap_top = []
for col_idx in sorted_idx:
fname = self.feature_names[col_idx] if col_idx < len(self.feature_names) else f"feat_{col_idx}"
shap_top.append((fname, float(vals[col_idx])))
result["shap_features"].append({
"feature": fname,
"shap_value": float(vals[col_idx]),
"value": float(x_clean[col_idx]), # correct column index
})
# Phishing risk features: positive SHAP (push toward phishing)
result["phishing_risk_features"] = [
{"feature": f, "shap_value": v}
for f, v in sorted(shap_top, key=lambda t: t[1], reverse=True)
if v > 0
][:top_n]
except Exception as exc:
log.warning(f"SHAP explanation failed: {exc}")
# --- LIME -----------------------------------------------------------
lime_top: List[Tuple[str, float]] = []
if self._lime_explainer is not None:
try:
lime_exp = self._lime_explainer.explain_instance(
x_clean,
self.model.predict_proba,
num_features=top_n,
labels=(1,), # Explain phishing class
)
lime_top = lime_exp.as_list(label=1)
# Build a name->column_index lookup for correct value retrieval
_name_to_col: dict = {}
if self.feature_names:
_name_to_col = {n: i for i, n in enumerate(self.feature_names)}
lime_features_out = []
for feat, weight in lime_top:
# Extract base feature name from LIME condition string
# e.g. "url_domain_length > 0.50" -> "url_domain_length"
base_name = feat
for op in (" <= ", " > ", " < ", " >= ", " = "):
if op in feat:
base_name = feat.split(op)[0].strip()
break
col_idx = _name_to_col.get(base_name)
feat_val = float(x_clean[col_idx]) if col_idx is not None else 0.0
lime_features_out.append({
"feature": feat,
"weight": float(weight),
"value": feat_val,
})
result["lime_features"] = lime_features_out
except Exception as exc:
log.warning(f"LIME explanation failed: {exc}")
# --- Agreement score ------------------------------------------------
result["agreement_score"] = self._compute_agreement(shap_top, lime_top, top_n=5)
return result
def _compute_agreement(
self,
shap_top: List[Tuple[str, float]],
lime_top: List[Tuple[str, float]],
top_n: int = 5,
) -> float:
"""Compute Jaccard similarity between top-N SHAP and LIME features.
Agreement score = |SHAP_top ∩ LIME_top| / |SHAP_top ∪ LIME_top|
Args:
shap_top: SHAP top features (name, value) tuples.
lime_top: LIME top features (condition, weight) tuples.
top_n: Top-N features to compare.
Returns:
Float in [0, 1]. 1.0 = perfect agreement.
"""
if not shap_top or not lime_top:
return 0.0
# Extract base feature names from LIME (conditions include comparison operators)
def _extract_name(lime_feat: str) -> str:
# LIME produces e.g. "url_domain_length > 0.50" — extract base name
for op in [" <= ", " > ", " < ", " >= ", " = "]:
if op in lime_feat:
return lime_feat.split(op)[0].strip()
return lime_feat.strip()
shap_names = set(name for name, _ in shap_top[:top_n])
lime_names = set(_extract_name(feat) for feat, _ in lime_top[:top_n])
intersection = len(shap_names & lime_names)
union = len(shap_names | lime_names)
return float(intersection / union) if union > 0 else 0.0
def batch_shap_values(self, X: np.ndarray) -> np.ndarray:
"""Compute SHAP values for a batch of emails.
Used for population-level feature importance analysis and
generating SHAP summary plots.
Args:
X: Feature matrix shape [n_samples, n_features].
Returns:
SHAP values array shape [n_samples, n_features].
"""
if self._shap_explainer is None:
return np.zeros_like(X)
X_clean = np.nan_to_num(X.astype(np.float32), nan=0.0, posinf=0.0, neginf=0.0)
try:
vals = self._shap_explainer.shap_values(X_clean)
if isinstance(vals, list):
return vals[1] # Return phishing class values
return vals
except Exception as exc:
log.warning(f"Batch SHAP failed: {exc}")
return np.zeros_like(X_clean)
|