MiLTL: A Cross-Modal Contradiction Cascade for On-Device Voice Phishing Detection

Abstract

Voice phishing (vishing) causes financial and psychological harm, yet deployed detectors that scan transcripts for scam vocabulary are more fragile than their benchmark scores suggest: their near-perfect accuracy on standard Korean corpora is memorization, collapsing under scammer paraphrase or language-model rewriting. We construct KorMMP, a benchmark keeping real regulator-sourced scam audio and real benign speech but decorrelating transcript from label, and MiLTL, an on-device detector built on affective and neutrosophic channels, including a cross-modal contradiction signal (XM) formulated to weigh lexical warmth against vocal coldness. In the reported evaluation, XM’s measured contribution is confidence banding and escalation routing rather than ranking accuracy. A scoring rule with zero gradient-learned parameters at inference screens every segment, referring ambiguous calls to a small on-device language model; raw audio stays on the device. We measure both stages on a commodity CPU container and smartphone; these budgets cover the detector only and exclude speech recognition, which we expect to dominate a live end-to-end budget. With no training on the hard benchmark, MiLTL achieves an AUROC of 0.965 (recording-level cluster bootstrap 95% CI 0.943–0.984), while every evaluated comparator stays below 0.69: text-only, audio-only, fusion, 7B multimodal. MiLTL scores lower on the saturated corpus, as expected for a detector designed not to rely primarily on lexical shortcuts. KorMMP is a controlled stress test of lexical decorrelation, not a measure of in-the-wild detection; its harmful and benign audio come from different corpora, so the source and label are structurally confounded, and the residual source effects cannot be fully excluded. XM is author-defined, constructed rather than naturally occurring in the synthetic stratum and not yet validated against human perception; this is the principal open limitation of the work. Within that scope, the work provides a reproducible basis for on-device vishing defense and a benchmark for detector robustness under vocabulary shift.