# Financial Fraud Prevention using AI Topic radar · 2,122 documents · 2021-01-01–2026-10-22 · updated 2026-10-30 Canonical: https://slate.greyb.com/radar/financial-fraud-prevention-using-ai/innovations ## In one minute ### The decision now runs inside the transaction The authorisation window sets the number the filings compete on. One payment design reports under 200 milliseconds using a tree-based scorer plus a layer watching behaviour over short, medium and long horizons [1]. Another reports under 50 milliseconds by running an autoencoder and a clustering check on devices close to the bank [2]. The same shift shows outside cards and transfers. A claims engine approves simple claims instantly and routes complex ones to a person [3]. A toll gate refuses a mismatched vehicle at the barrier [4]. A smart contract permits, delays or blocks a transfer on a score delivered by an oracle [5]. Instant rails make this a structural requirement: a payment that cannot be recalled has to be judged before settlement, so the score competes with the authorisation itself. 1. [IN202621052344A](/?paper=IN202621052344A&utm_source=radar&utm_medium=dashboard) 2. [IN202621064237A](/?paper=IN202621064237A&utm_source=radar&utm_medium=dashboard) 3. [IN202541103971A](/?paper=IN202541103971A&utm_source=radar&utm_medium=dashboard) 4. [IN202541126229A](/?paper=IN202541126229A&utm_source=radar&utm_medium=dashboard) 5. [IN202531088117A](/?paper=IN202531088117A&utm_source=radar&utm_medium=dashboard) ### The largest sums bought the authenticated session, where the patents are thinnest Visa agreed in August 2026 to buy BioCatch for $2.4 billion in cash, with completion expected by the end of its 2027 fiscal second quarter. BioCatch scores keystrokes and touch gestures continuously on 1.8 billion devices across roughly 350 banks, and the stated purpose is to catch scams and account takeover before a payment is reached. The deal follows Visa's $946 million purchase of Featurespace, a behavioural analytics firm. ### Company output is growing, and the academic tail is growing faster Patent publications in the record ran 105 in 2021 and about 730 in the ten months to 30 October 2026, with 2026 a partial year. Companies still publish the most in every year: 77 in 2021, 257 in 2025, 308 in 2026. What changed is the rest of the field. ### Four stages appear in every area Read the filings as one pipeline and the areas line up. Signals are fused, a model scores, the score maps to an action, the action is recorded, and a monitor watches for the model going stale. What differs between areas is the first box and the third. Insurance adds a payout decision to the score. Identity verification turns the score into a choice of which check to run. Toll and on-chain work turn it into a physical or contractual block, which is a harder action than a flag and a much harder thing to reverse. ### Where results are published, the gain lands in false positives first Bank of Ireland is the one institution in this record with a measured outcome. Since deploying Featurespace's real-time platform in 2024 it reports a 25 percent fall in customer payment fraud losses, a 30 percent fall in attempted fraud, an 87.1 percent fall in false positives and an 85 percent fall in alert volumes, with case handling time halved. Celent reviewed the programme and describes the same picture as a shift from batch review to real-time decisioning. ### The operating filings sit with about a dozen institutions The geographic split is blunt. Chinese companies hold 435 patent documents and US companies 372, and those two groups contain the card networks, processors, banks and scoring vendors. Indian companies hold 39, against 410 from Indian individuals and 345 from Indian academics. ### Three model families recur, and the newest filings fuse them The default is a supervised classifier for known schemes paired with an anomaly or reconstruction component for new ones, on the reasoning that a rule can only catch what someone wrote down in advance. Relational models score an entity by its neighbours and have become the second standard, applied to account graphs and identity graphs joined across institutions. Sequence models handle anything that unfolds over time. ### Supervisors now write the data-sharing arrangement into the rule India's insurance regulator issued a fraud framework taking effect 1 April 2026 that requires a board-approved anti-fraud policy, a dedicated monitoring committee, and mandatory participation in the Insurance Information Bureau's fraud monitoring technology framework and its caution repository of blacklisted parties. The repository uses a unique identifier to follow a policyholder across insurers, and it is a condition of holding a licence rather than a product choice. A public insurance registry that could attach a score to a policyholder is under development, which extends the same logic into underwriting. ## Overview A fraud decision that has to complete before settlement is bounded by three things, and the radar's evidence is organised around them. The first is time: filings compete on the authorisation window, with designs reporting under 200 milliseconds and under 50 milliseconds. ## Research ### The graph record is larger than the evidence behind it Graph work appears in five of the six areas, and the filings assume it beats tabular models. A 2026 benchmark tested that across ten fraud datasets and found tree ensembles outperformed their graph counterparts in most cases, with graph models doing relatively worse as labels became scarce. ### The patented invention has moved to the maintenance loop This is the clearest technical change in the recent record. Actimize watches transaction features for sudden and gradual shifts, weights them by importance, and updates only when the weighted score clears a threshold. Fair Isaac measures concept drift against features generated per account type and corrects the live score without retraining. Visa compares feature-similarity patterns between live traffic and training history and flags a mismatch without ground-truth labels. ### Labels limit what these models can learn Getting an outcome back takes weeks, and the filings compete on handling that delay. Equifax uses older balanced data to relabel recent transactions while chargebacks are still arriving. Mastercard expands training with high-confidence predictions on unlabelled payments and filters declined transactions so false fraud labels do not corrupt retraining. Visa uses reinforcement learning to re-label noisy training records before the detector is trained. Capital One splits uncertainty into data noise and model ignorance, self-correcting for the first and sending the second to a human. ### Sharing is the same architecture with a contested seam Every institution trains locally and sends a transformed model state to a hub; what crosses is argued over. Gradients keep records at home and still carry information about them. Synthetic transactions move the risk to whether the fake data preserved the patterns worth sharing. Masked updates treat the hub as an adversary. ### Explanation is now an output state Explanation has moved from the write-up into the decision. One filing maps a calibrated score to five outcomes, from authorise through step-up authentication to block, with an uncertainty value attached to each. Another states that the system explains why a transaction was flagged so a bank can verify rather than decline. Toll and cross-chain systems emit topological evidence chains and log each classification to a permissioned ledger. ### The most valuable input belongs to someone else The score improves with reference data the institution does not own. Carrier risk tags arrive under secure computation, government registries supply the identity check, and payer-supplied reference signatures decide whether matching runs in its strong or weak mode. A model is only as good as the data-sharing arrangement behind its inputs, which puts negotiating power with whoever holds the reference rather than whoever writes the classifier. ### No filing reports a measured production result Every design claim in this record is architectural: what is fused, how often it updates, what state the decision maps to. No filing reports a detection rate, a false-positive rate, a loss reduction or a deployment outcome. The measured numbers in the field come from academic work on public or simulated datasets, which is a different setting from a live authorisation stream. ## Market ### The supply side is consolidating onto the rail, and the standalone cloud scorer has left Amazon closed Fraud Detector to new customers on 7 November 2025 and has kept it in maintenance with no new features since. A general-purpose cloud scoring service, priced on API calls and indifferent to the payment rail, is the shape that lost. ## Strategy ### Bank of America moved position fastest, and it did so by widening its surface Bank of America went from a minor filer to 23 patent documents in 2025 and 2026, level with ICBC at the top of the field over that period. Its recent filings spread across channels instead of deepening one technique. ### The consortium is turning into the unit of work The cross-border and federated area is the smallest at 23 documents, and the shape of its growth is the signal: one publication in 2021, three in 2025, eleven in 2026. Bank of America's filings there assume institutions keep their records and share only a transformed model state, and its kiosk work applies the same logic to a shared minimum standard across member banks. ### Specialists hold narrow positions while institutions hold the surfaces The filing pattern separates three kinds of participant, and each sits in a different part of the record. ### Two areas carry dense filing activity and thin evidence of a market Emerging and specialised fraud published 151 patent documents in 2026. Blockchain and telecom filings there come mostly from institutes and individuals, and the commercial positions are narrow and product-shaped. Insurance shows the same asymmetry in a different form: 48 filings in 2026, while the evidence that claims fraud is expensive sits with incumbent vendors rather than with these filers. ### Deployment evidence comes from networks and supervisors, and almost never from a vendor Where a live result appears, the operator is a coalition, a network or a central bank: the Taiwanese consortium above, Swift's cross-bank trial, Singapore's pooled data exercise, India's intelligence-sharing agreement. ## Areas - Transaction & Payment Fraud · 1,146 documents - Credit & Loan Fraud · 280 documents - Insurance & Claim Fraud · 153 documents - Identity & Account Fraud · 65 documents - Cross-Border & Federated Fraud · 23 documents - Emerging & Specialized Fraud · 405 documents