File size: 2,157 Bytes
cf9e03f | 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 | """
Reinforced Disentangled HTML/DOM Representation Learning with Hard-Sample Mining
(Yoon, Buu, & Kim, 2025)
Separates structural DOM tree topology from surface text semantics to expose obfuscated phishing templates.
"""
import numpy as np
class DisentangledDOMEncoder:
"""
Projects raw HTML/DOM tree metrics into two orthogonal subspaces:
1. Structural Subspace (z_struct): DOM depth, tag transitions, hidden form/iframe manipulations.
2. Semantic Subspace (z_sem): Textual token distribution and form field labels.
Includes Hard-Sample Mining loss weighting to focus gradient updates on deceptive clones.
"""
def __init__(self, input_dim: int = 24, latent_dim: int = 32, seed: int = 101):
rng = np.random.default_rng(seed)
self.W_struct = rng.normal(0, np.sqrt(2.0 / input_dim), size=(input_dim, latent_dim))
self.b_struct = np.zeros(latent_dim)
self.W_sem = rng.normal(0, np.sqrt(2.0 / input_dim), size=(input_dim, latent_dim))
self.b_sem = np.zeros(latent_dim)
def forward(self, x_dom: np.ndarray) -> tuple[np.ndarray, np.ndarray, float]:
"""
Computes disentangled structural and semantic representations and their orthogonality penalty.
"""
z_struct = np.tanh(x_dom @ self.W_struct + self.b_struct)
z_sem = np.tanh(x_dom @ self.W_sem + self.b_sem)
# Orthogonality regularization loss: || Z_struct^T @ Z_sem ||_F^2
corr_matrix = (z_struct.T @ z_struct) if z_struct.shape[0] == 0 else (z_struct.T @ z_sem) / max(1, x_dom.shape[0])
ortho_loss = float(np.mean(corr_matrix ** 2))
z_combined = np.concatenate([z_struct, z_sem], axis=-1)
return z_combined, z_struct, ortho_loss
@staticmethod
def mine_hard_samples(probs: np.ndarray, labels: np.ndarray, margin: float = 0.25) -> np.ndarray:
"""
Hard-Sample Mining (Yoon et al., 2025): Assigns amplified sample weights (2.5x)
to borderline or misclassified samples where |y - p| > margin.
"""
errors = np.abs(labels - probs)
weights = np.where(errors >= margin, 2.5, 1.0)
return weights
|