X-MAP: eXplainable Misclassification Analysis and Profiling for Spam and Phishing Detection

Qi Zhang; Dian Chen; Lance M. Kaplan; Audun Jøsang; Dong Hyun Jeong; Feng Chen; Jin-Hee Cho

cs.AIFeb 17, 2026

X-MAP: eXplainable Misclassification Analysis and Profiling for Spam and Phishing Detection

Authors: Qi Zhang, Dian Chen, Lance M. Kaplan, Audun Jøsang, Dong Hyun Jeong, Feng Chen, Jin-Hee Cho

Organizations: Virginia Tech, USA · U.S. Army DEVCOM Army Research Laboratory, USA · University of Oslo, Norway · University of the District of Columbia, USA ·

Abstract

Misclassifications in spam and phishing detection are very harmful, as false negatives expose users to attacks while false positives degrade trust. Existing uncertainty-based detectors can flag potential errors, but possibly be deceived and offer limited interpretability. This paper presents X-MAP, an eXplainable Misclassification Analysis and Profilling framework that reveals topic-level semantic patterns behind model failures. X-MAP combines SHAP-based feature attributions with non-negative matrix factorization to build interpretable topic profiles for reliably classified spam/phishing and legitimate messages, and measures each message's deviation from these profiles using Jensen-Shannon divergence. Experiments on SMS and phishing datasets show that misclassified messages exhibit at least two times larger divergence than correctly classified ones. As a detector, X-MAP achieves up to 0.98 AUROC and lowers the false-rejection rate at 95% TRR to 0.089 on positive predictions. When used as a repair layer on base detectors, it recovers up to 97% of falsely rejected correct predictions with moderate leakage. These results demonstrate X-MAP's effectiveness and interpretability for improving spam and phishing detection.

Figures & tables

Figure 1: X-MAP for explainable misclassification analysis in spam/phishing detection.

Figure 2: X-MAP Re-evaluation for Error Recovery: This enables recovery of correctly classified samples while keeping error leakage low.

CFR	GP type	TP $\to$ TP GP	FP $\to$ TP GP	TN $\to$ TN GP	FN $\to$ TN GP
LogReg	Ori.	0.2051 $\pm$ 0.1405	0.5490 $\pm$ 0.1440	0.0543 $\pm$ 0.0524	0.1213 $\pm$ 0.0518
	Vac.	0.1946 $\pm$ 0.1281	0.5010 $\pm$ 0.1180	0.0536 $\pm$ 0.0500	0.1207 $\pm$ 0.0521
	Diss.	0.2213 $\pm$ 0.1320	0.5336 $\pm$ 0.1164	0.0603 $\pm$ 0.0634	0.1467 $\pm$ 0.0658
	Ale.	0.0093 $\pm$ 0.0093	0.0600 $\pm$ 0.0338	0.0090 $\pm$ 0.0136	0.0271 $\pm$ 0.0149
	D ${\alpha}$	0.0050 $\pm$ 0.0050	0.0202 $\pm$ 0.0125	0.0006 $\pm$ 0.0014	0.0015 $\pm$ 0.0011
	D ${\beta}$	0.0015 $\pm$ 0.0017	0.0105 $\pm$ 0.0115	0.0001 $\pm$ 0.0003	0.0003 $\pm$ 0.0003

Table 1: Jensen–Shannon divergence between individual messages (IM) and group topic profiles (GP) on the SMS dataset

CFR	Metric	Topic Representation								UQ-D
CFR	Metric	Ori.	Ale.	Vac.	Diss.	REL-U	$D_{\alpha}$	$D_{\beta}$	ODIN	UQ-D
LogReg	AUROC (P)	0.9409	0.9763	0.9470	0.9474	0.8283	0.8954	0.8545	0.8545	0.9367
	FRR@95%TRR (P)	0.3988	0.0895	0.3923	0.2301	0.2321	0.5900	0.6054	0.6054	0.6304
	AUROC (N)	0.8807	0.8253	0.8802	0.8711	0.5252	0.8536	0.8041	0.8041	0.9369
	FRR@95%TRR (N)	0.3588	0.6186	0.3574	0.3971	0.9397	0.4130	0.5515	0.5515	0.2247
SVM	AUROC (P)	0.9574	0.9264	0.8570	0.9063	0.9057	0.9316	0.9090	0.9090	0.9734

Table 2: Performance of X-MAP on SMS Dataset

CFR	Metric	Topic Representation
CFR	Metric	Ori.	Ale.	Vac.	Diss.	REL-U	$D_{\alpha}$	$D_{\beta}$	ODIN
LogReg	AUROC (P)	0.9504	0.9523	0.9470	0.9474	0.8283	0.9728	0.8545	0.8545
	FRR@95%TRR (P)	0.3066	0.1679	0.3923	0.2301	0.2321	0.2162	0.6054	0.6054
	AUROC (N)	0.9471	0.9615	0.8802	0.8711	0.5252	0.8771	0.8041	0.8041
	FRR@95%TRR (N)	1.0000	1.0000	0.3574	0.3971	0.9397	0.3905	0.5515	0.5515
SVM	AUROC (P)	0.9454	0.9000	0.9143	0.9184	0.9251	0.9471	0.9049	0.9049

Table 3: Performance of X-MAP on Messages Falsely Rejected by UQ-Based Detector

CFR	Metric	Topic Representation
CFR	Metric	Ori.	Ale.	Vac.	Diss.	REL-U	$D_{\alpha}$	$D_{\beta}$	ODIN
LogReg	RecovR	0.9436	0.9498	0.9404	0.9279	0.1923	0.2268	0.8694	0.8276
	# Recovery	903	909	900	888	184	217	832	792
	LeakR	0.1542	0.1522	0.1542	0.1522	0.1502	0.1502	0.1542	0.1502
	# Leakage	78	77	78	77	76	76	78	76
	# Correct Fix	825	832	822	811	108	141	754	716

Table 4: Performance of X-MAP on Messages Rejected by UQ-Based Detector

Phishing Detection

See all 16 papers

AURA: Adaptive Uncertainty-Routed Analysis for Email Threat Detection
Sep 17, 2026Omran Berjawi, Walid fahs, Rida KhatounDistribution Shift Robustness Phishing Detection
Institut Polytechnique de Paris, Télécom Paris, Palaiseau, France · Islamic University of Lebanon, Faculty of Engineering, Wardanieh, Lebanon
Robustness, Cost, and Attack-Surface Concentration in Phishing Detection
Mar 19, 2026Julian Allagan, Mohamed Elbakary, Zohreh Safari +4Phishing Detection Adversarial Robustness
Department of Mathematics, Computer Science, and Engineering Technology Elizabeth City State University, Elizabeth City, NC 27909, USA
A Hybrid, Multi-Layered Pipeline for Phishing and Threat Classification: Independently Validated URL and NLP Engines with a Calibrated Multi-Channel Fusion Stage
Jun 19, 2026Saifelden M. Ismail, Aser O. Ibrahim, Omar A. MahmoudPhishing Detection Multimodal Fusion
Department of Communications and Information Engineering Zewail City of Science and Technology, Giza, Egypt

Sep 17, 2026J/K move · Enter open · S save

AURA: Adaptive Uncertainty-Routed Analysis for Email Threat Detection

Omran Berjawi, Walid fahs, Rida Khatoun

Institut Polytechnique de Paris, Télécom Paris, Palaiseau, France · Islamic University of Lebanon, Faculty of Engineering, Wardanieh, Lebanon

Distribution Shift Robustness Phishing Detection

Open paper AskPDF arXiv

Email spam and phishing attacks remain a critical security threat. Adversaries increasingly exploit large language models to craft contextually convincing malicious messages, and existing spam detection systems often struggle to keep pace. Generalization across diverse and evolving attack scenarios is limited, which reduces effectiveness once these systems are deployed in practice. This paper introduces Adaptive Uncertainty-Routed Analysis (AURA), a multimodal email threat detection system that analyzes both the content of an email and its embedded URLs. AURA is built around two layers: the first quantifies prediction uncertainty from a URL classifier, and only ambiguous messages are escalated to a fine-tuned transformer encoder for semantic analysis. The system is evaluated on eight heterogeneous training corpora together with two held-out real-world corpora spanning a decade of adversarial campaigns. AURA reaches a macro F1-score of 0.9858 in-distribution, and on NazPhish-Eval and GuenterTrap-Eval it maintains 0.9502 and 0.9436, respectively, which is evidence of robust generalization under genuine distribution shift.

Abstract

Phishing Detection

Explore similar work

AURA: Adaptive Uncertainty-Routed Analysis for Email Threat Detection

Robustness, Cost, and Attack-Surface Concentration in Phishing Detection

A Hybrid, Multi-Layered Pipeline for Phishing and Threat Classification: Independently Validated URL and NLP Engines with a Calibrated Multi-Channel Fusion Stage