Safe digital finance under constraints

Account ownership among Filipino adults rose from 29% in 2019 to 56% in 2021, driven by e-money accounts, which went from 8% to 36% of adults over the same period (BSP, 2021 Financial Inclusion Survey). The 2025 Global Findex still places the Philippines at about 50% account ownership, below the East Asia and Pacific average (World Bank, 2025). Many of these account holders are new to formal finance, and they are heavily targeted: in a 2025 survey of 1,000 Filipinos, 77% had encountered a scam in the previous year and 31% had lost money to one (Global Anti-Scam Alliance, State of Scams in the Philippines 2025). Complaints to the central bank reached about 70,000 in 2024, of which roughly 13% concerned unauthorized transactions such as phishing (BusinessMirror, citing BSP, June 2025). Uncontrolled fraud is known to erode the consumer benefit and the inclusion gains of mobile financial services (Buku and Mazer, CGAP, 2017), and predatory loan, gambling, and trading apps exploit users who cannot read what they consent to: in one study of such apps in an emerging market, 85% of participants did not understand basic app permissions (Pervez et al., 2026).

This program therefore treats protection as a precondition for inclusion, and treats the constraints of the market (low bandwidth, low-end devices, informality) as design variables rather than obstacles to be assumed away.

Recent work suggests that useful scam detection no longer requires a large model or a server. A 0.5-billion-parameter model trained by distillation reached 94% accuracy on SMS threat detection (ElZemity et al., 2026), and a comparison of fifteen open models for phishing-website detection found that 1-billion-parameter models answer in under a second where 70-billion-parameter models take 23 to 31 seconds, at a cost in accuracy that the study quantifies (Goldenits et al., 2025). Detectors are also brittle against obfuscated and code-switched text in low-resource settings, which is the Philippine case (Adversarial robustness in smishing detection, 2026). Whether any of this holds for Philippine consumer-finance messages and terms, in English and Taglish, on a low-end phone, is an open question.

The intended first study would be a benchmark of consumer-finance cases built from public Philippine sources, such as regulator advisories and published e-wallet and lending terms, with answer keys stating the harm and the safe next step. It would compare frontier and small on-device models, each released directly and through deterministic checks, on correctness, serious failures released, useful answers withheld, cost, and the human verification time that remains. The method would follow the program’s sister study on assessment support in higher education.

The program is in scoping. No detector, score, or ranking is published, and no claim is made that any system protects users. Alternative credit scoring, household resilience measurement, and open-finance governance are adjacent literature, not outputs; the longer trajectory toward a resilient ecosystem that can support open finance is noted only as direction.