Digital Wallets: Security Models Compared
Not all digital wallets store money the same way. A comparison of custodial, bank-linked, tokenized card, and self-custody models — and what each means for fraud, recovery, and privacy.
By Velloreth
Say “digital wallet” and most people picture one thing: an app where money sits. That picture is wrong in an interesting way. The apps we call wallets sit on at least four fundamentally different security models, with different custody, different authentication, and — most importantly — different answers to the question that actually matters after something goes wrong: who is liable, and what can you do about it?
This post compares the four models directly, because the security differences between them are larger than the differences between any two wallet brands.
The four models
1. Custodial e-money wallets
The model behind most South and Southeast Asian wallets — eSewa, Khalti, GrabPay, and their peers — plus anything the regulator classifies as electronic money.
How it works: You load money into a balance held by the licensed operator (or a partner bank ring-fenced from the operator’s own funds). Spending debits that balance; the operator moves money to merchants and other users on its internal ledger, and settles with banks and payment systems in the background. Regulation matters here: Nepal Rastra Bank licenses payment service providers and imposes e-money safeguards and KYC tier rules; the EU’s EMD2 requires safeguarding of customer funds. The specifics differ by market, but the structure is consistent: your balance is a liability of a regulated non-bank entity.
Security properties:
| Aspect | How it works | Weak point |
|---|---|---|
| Fund custody | Operator-held, usually ring-fenced | Operator insolvency, or outright fraud by the operator |
| Authentication | App PIN + OTP on critical actions | SIM-swap attacks against the OTP channel |
| Device binding | Common, improving | Device swap plus social engineering of support |
| Fraud liability | Set by regulation and operator policy | Recovery often slow; disputed push transfers hard to reverse |
| Recovery | Operator support with KYC re-verification | Support desks themselves are a major attack surface |
The honest assessment. Custodial wallets gave the unbanked a working digital account, and that is a real achievement. Their security model concentrates trust in one company’s systems and one company’s customer support. The most common serious failures are not cryptographic — they are SIM swaps intercepting OTPs, phishing that harvests the app PIN, and social engineering of the support process that resets access. Funds sitting idle in a large balance also create a target that a bank deposit insurance regime does not cover.
2. Bank-linked wallets and rails
The model behind UPI apps in India, pay-by-bank wallets, and anything that spends directly from your bank account without storing a balance.
How it works: The wallet app holds no money. It holds your authorization to initiate debits from your bank account, authenticated through the payment rail — for UPI, a PIN set at the rail level, with the bank and the app each enforcing their own controls. Nepal’s connectIPS-based wallets and mobile banking apps follow the same pattern: the balance lives at the bank; the app is an authorization surface.
Security properties:
| Aspect | How it works | Weak point |
|---|---|---|
| Fund custody | Your bank, under normal deposit rules | None specific — this is the mainstream banking model |
| Authentication | App lock + rail PIN + OTP for setup | Push-payment fraud: you authorize the transfer yourself |
| Device binding | Increasingly enforced on UPI and similar | Account takeover at the app level still moves money |
| Fraud liability | Regulator-set; push payments often customer’s risk | Real-time means irreversible; clawback depends on bank policy |
| Recovery | Bank fraud desk; NRB/RBI complaint channels | Slower than people expect for “instant” payments |
The honest assessment. No stored balance means no float for an operator to lose and no wallet insolvency risk — the money is either in your bank or it is gone. The trade-off is that the dominant fraud vector becomes you: authorized push payments, where a convincing scam gets the victim to approve the transfer themselves. Real-time rails were designed for speed, not reversibility, and every regulator in the region has spent the last few years writing rules around exactly that tension — cooling-off periods for new payees, transaction limits, and liability frameworks that increasingly put partial responsibility on banks and apps rather than customers, without eliminating customer responsibility entirely.
3. Tokenized card wallets
Apple Pay, Google Pay, and anything that stores your card as a token rather than a number.
How it works: When you add a card, your bank or card network issues a device-specific token — a substitute card number that works only in combination with that device. Each transaction generates a one-time cryptogram, authorized by your device’s secure element and your biometric or device PIN. The merchant never sees your real card number; a merchant-side breach yields tokens that are useless elsewhere.
This is the security model most people interact with without knowing it exists, and it is genuinely the strongest conventional-payments design of the last two decades. Network tokenization extends the same idea beyond device wallets to stored card credentials in e-commerce, replacing static card numbers with tokens that can be rotated instantly if compromised.
Security properties:
| Aspect | How it works | Weak point |
|---|---|---|
| Credential exposure | Real card number never leaves bank/network | Initial enrollment (card entry) remains a phishable moment |
| Per-transaction auth | Biometric + cryptogram | None significant at the transaction layer |
| Breach containment | Tokens are device- and context-bound | Merchant breaches of untokenized fallback paths still hurt |
| Fraud liability | Card network chargeback rules apply | Strongest consumer protections of any model here |
| Recovery | Card reissue through bank; token revocation is instant | Losing the device is manageable by design |
The honest assessment. Tokenized card wallets have the best liability story (card network chargeback protections), the best per-transaction security (biometric plus cryptogram), and the best breach containment (revocable, device-bound tokens). Their weakness is upstream: getting the card into the wallet still involves typing a number into a phishable field, and the underlying card network’s economics (interchange, cross-border fees) remain unchanged. In markets where cards barely exist, this model is irrelevant regardless of how good it is.
4. Self-custody wallets
Crypto wallets — MetaMask, hardware wallets, anything where you hold the keys.
How it works: You hold a private key (usually represented as a seed phrase). Every transaction is cryptographically signed by you. No operator, no bank, no issuer: the ledger itself is the bank. What you lose in recourse you gain in censorship resistance and portability.
Security properties:
| Aspect | How it works | Weak point |
|---|---|---|
| Fund custody | You, via key material | Key loss = permanent fund loss |
| Authentication | Device PIN, biometric unlock, signing approval | Phishing for signing approvals; drainer contracts |
| Attack surface | Software wallet extensions, malware | Browser extensions are a proven catastrophic risk |
| Fraud liability | You | No chargebacks, no support desk, no reversal |
| Recovery | Pre-arranged: multisig, social recovery, inheritance plans | Most users set up none of these |
The honest assessment. Self-custody is a security model for people willing to be their own security team. For that audience it works: hardware signing, multisig vaults, and pre-arranged recovery produce a level of asset control no custodian offers. For everyone else, it converts every mistake into a total loss with no appeals process. The industry’s repeated discovery that most humans will not securely store a 12-word phrase is not a temporary adoption gap; it is a permanent design constraint.
The fifth thing people mistake for a model: crypto inside custodial wallets
Many mainstream wallet apps have added crypto buying and holding. This is custodial e-money’s security model applied to a volatile asset — with the operator’s custody risk, plus market risk, plus whatever the local regulator does or does not cover. Nepal’s position on cryptocurrency trading has been restrictive, with NRB warnings against crypto transactions for payment purposes; residents of such markets holding crypto through any channel are operating in regulatory gray zones that no amount of app polish removes.
Where tokenization is quietly spreading
The most interesting recent development is tokenization escaping the card world. Account-based rails that historically transmitted real identifiers — bank account numbers, phone numbers used as proxies, virtual payment addresses — are adopting token layers. NCHL’s connectIPS tokenization work is a concrete example: replacing sensitive bank details with unique tokens for recurring e-mandates, wallet loads, and e-commerce checkouts, with early adoption figures in the tens of thousands of users and hundreds of thousands of transactions within the first months of rollout.
The mechanism mirrors network tokenization on cards: the application never handles the real account identifier, the token is revocable without changing the underlying account, and a breach of the application layer yields tokens that are useless outside their defined context. Applying it to request-to-pay and e-mandate flows is the right place to start, because those flows store credentials for reuse — exactly where static identifiers have historically leaked.
If you evaluate a wallet today, one useful question is whether it tokenizes stored payment credentials at all. Wallets that still store and transmit raw account or card identifiers for recurring payments are carrying 2010s risk in a 2020s product.
Choosing a wallet by the questions that matter
Brand comparisons are mostly noise. When you evaluate a wallet — for yourself or as a product decision — these are the questions with actual security consequences:
Where does my money sit when I am not spending it? Custodial balance, bank account, card network, or on-chain. Each has its own failure mode. Money you cannot afford to lose belongs wherever the failure mode is most boring.
What does it take to move money out? The weakest authentication step on the critical path defines your real security, not the strongest one advertised. If a device change plus a phone call to support can drain the account, that is your security model.
Who eats fraudulent push payments? Read the liability policy, not the marketing. Card-network wallets have chargeback regimes; custodial wallets have operator policies that vary; bank-linked push payments fall under whatever regulatory framework your market has written — some markets have shifted partial liability to institutions in recent years, others have not.
How does recovery work before I need it? Test the mental path: lost phone, SIM changed, email compromised. If recovery means calling a support desk that will social-engineer-reset you into the attacker’s hands, you have found the real attack surface. Hardware security keys for the email controlling financial apps are the single highest-leverage upgrade most people skip.
Is the balance insured or safeguarded? These are different words with different meanings. Safeguarded e-money means ring-fenced, not lent out. Insured deposits mean a government-backed guarantee up to a limit. Knowing which one you have tells you what a bank or operator failure costs you.
The practical bottom line
For daily spending in regulated markets, a bank-linked wallet with strong device binding, or a tokenized card wallet, has the best security properties per unit of convenience. For money that needs to survive an operator’s bad day, a bank account with deposit insurance beats any wallet balance. Self-custody has a genuine place for people who understand key management and want sovereign control, and almost nowhere else.
The wallet apps themselves have converged on the surface — PIN, biometric, clean UI — while the models underneath have diverged more than ever. Pick the model first. The app is the easy part.