釣魚攻擊（Phishing Attack）藉由偽造合法網站統一資源定位符（URL）誘導使用者洩漏帳號密碼與財務資訊，防禦攻擊的核心困境在於黑名單更新速度長期落後於攻擊者輪替網域的速度，使以靜態特徵拼接為主的現有偵測方法在灰色邊界案例中存在系統性誤判，且多數系統缺乏可供資安人員快速複核的決策依據。本研究設計並實作 GateNet，採六層推理架構，核心為 Softmax 正規化閘控多層感知機（Gate MLP），針對每一條輸入 URL 個別計算三路特徵權重，對統計特徵（Stat，41 維）、字元卷積神經網路特徵（Char-CNN，128 維）及語意閘控循環單元特徵（Semantic GRU，128 維）執行融合加權，閘控輸出直接作為可解釋的特徵重要性指標。架構中額外設計領域感知校準層（Step 2a），透過 107 筆可信域名白名單與三層 Typosquatting 偵測機制（可疑 TLD、仿冒前綴、Levenshtein 編輯距離 ≤ 2）雙向校準模型輸出：對仿冒域名調高其釣魚機率以修正漏判，對可信域名則調低機率以避免誤判。並以閘控值審視，調整日誌與 Gemini 2.5 Flash 自然語言威脅報告構成三層可解釋輸出鏈；針對帶路徑良性 URL 的系統性誤判，另提出定向偏差修正方法，透過閘控分布日誌診斷訓練偏差並以針對性資料增強修正。於 587，638 筆修正資料集上，GateNet 測試集準確率達 99.82%、F1 分數 0.9940；逐階段效能自單分支（96.4%–99.7%）、靜態融合（99.79%）演進至動態閘控（99.82%），其改善經 McNemar 檢定具統計顯著性（p=0.003）。在 35 筆人工定向設計的模糊測試集（Fuzzy Test）上，涵蓋合法知名域名、仿冒域名、可疑灰色地帶與新創域名四類邊界情境，屬案例驗證性質，領域感知層將通過數由 20 提升至 32，修正全部 12 筆具可信域名錨點的漏判；其餘 3 筆無可信域名錨點的純通用詞釣魚 URL 為當前架構的結構性限制，完整案例清單與逐筆結果詳見附錄B。;Phishing attacks fabricate the Uniform Resource Locators (URLs) of legitimate websites to trick users into disclosing credentials and financial information. Their core difficulty is that blacklist updates persistently lag behind attackers′ domain rotation， causing detection methods based on static feature concatenation to misclassify gray-boundary cases， while most systems provide no decision rationale that security analysts can quickly review. This study designs and implements GateNet， a six-tier inference architecture whose core is a Softmax-normalized gated multilayer perceptron (Gate MLP). For each input URL it individually computes the weights of three feature streams (per-URL adaptivity) and performs weighted fusion over statistical features (Stat， 41-dim)， character-level convolutional features (Char-CNN， 128-dim)， and semantic recurrent features (Semantic GRU， 128-dim); the gate outputs serve directly as interpretable feature-importance indicators. A domain-aware calibration layer (Step 2a)， comprising a 107-entry brand whitelist and three-tier Typosquatting detection (suspicious TLD， spoofing prefix， Levenshtein distance ≤ 2)， bidirectionally calibrates the output: raising the phishing probability of spoofed domains to correct false negatives， and lowering it for legitimate brand domains to widen the safety margin. A three-tier interpretable output chain—gate inspection， adjustment log， and a Gemini 2.5 Flash threat report—is further provided， together with a targeted bias-correction method that diagnoses training bias from gate-distribution logs and applies targeted data augmentation. On a corrected dataset of 587，638 URLs， GateNet attains 99.82% test accuracy and an F1 score of 0.9940; accuracy progresses from single-branch models (96.4%–99.7%) through static fusion (99.79%) to dynamic gating (99.82%)， an improvement that is statistically significant under McNemar′s test (p=0.003). On a fuzzy-test set of 35 manually curated boundary cases spanning four scenarios (well-known legitimate domains， typosquatted domains， suspicious gray-zone URLs， and new/unknown domains)， the domain-aware layer raises the pass count from 20 to 32， correcting all 12 brand-anchored false negatives; the 3 remaining generic-word phishing URLs without brand anchors represent a structural limitation of the current architecture (the complete case list is provided in Appendix B).
