Customer care is an integral part of an organization. Customers contact care staff to perform many monetary and non-monetary actions. For example: New application/enrollment for a service or product Change address, password, add additional participants Balance inquiry, Payment, Transfers Block, Cancel account, subscriptions ... Many more...
Customer care professionals validate the identity of calling customers using security questions. Once identity is validated customers can perform monetary or non-monetary actions.
Fraudsters utilize many methods to collect customers details by hacking, dark web purchases, using malware, social engineering or theft. In cases of fraud, the detection will take a long time until customers dispute and
forensic analysis can be performed on calls, call recordings and all. In cases of fraud, organizations/customers will lose millions or billions of dollars. Few of the common fraudster actions in calls: Fraudulent purchases Account
Takeover - Change address and divert statements, checks Service disruption, cancellation, revenge Data theft... These fraudster actions can be reduced drastically if we enhance our user verification mechanism. Basic security questions
can't protect customers if their details are compromised or stolen, also Caller IDs can be spoofed etc.
InsightLake's call fraud detection solution enables customer care and fraud operations teams to detect call fraud in real time and monitor inbound calls using deep learning and machine learning models very effectively.
Call fraud solution uses following features to perform comprehensive call analysis. Like InsightLake's other solution "User Behavior Analytics" many of caller and call parameters are processed, baselines created, risk scores are calculated and link analysis with existing fraud database is performed. Caller profile is created and maintained. Call parameters - Caller ID, Network, VOIP Gateway, Carrier, User Agent, Contact ... from SIP, CDR etc Media parameters - Codec, buffers, packet loss, RTCP etc.. Audio modulation - Standardized voice sampling, DB Levels & Deep learning matching
Who is the caller - caller ids linked to an account Usually locations from where calls come Usually which devices caller calls from Usually which carriers, gateways calls comes from Caller voice signature - voice tone, standardized modulation with db levels, words list Caller natural language signature - Transcripts (Keywords modeling) Usual actions performed in past - password change, money transfer.. Historical actions Risk segmentation
Like our finger prints every person's voice has a unique signature. When we take voice modulation, SNR, db levels etc over few seconds of window a unique voice signature is generated. This signature provides sufficient data to our voice identification deep learning models to detect right customer profile vs fraudster.
Fraud solution includes complex machine learning based call risk scoring models. These models rely on call parameter's properties like reputation, weights, outlier buckets etc. Scoring is created from 0 to 100 range, 0 being the least risky call to 100 most risky. Along with risk scores thresholds are provisioned to bucket calls in Low, Medium and High buckets for operational analysis.
Fraud solution enables fraud ops team to maintain fraudulent caller database and performs link analysis with voice to detect if the caller is a known fraudster.
Fraud solution monitors every call, creates risk scores and capture relevant details for operations team to drill down risky calls.
Operations teams can search calls and drill down in details very easily using an intuitive interface.
Fraud solution provides interactive UI, which operations team can use to explore call details, look at risky calls in real time and perform actions like incident creation. REST integration is supported with many popular incident or case management solutions.
Fraud solution's goal is to detect call fraud without affecting customer experience for low risk callers. There are many scenarios which are handled effectively to reduce false positives. First Time Call - There are scenarios when there is no existing caller profile, in that case un supervised bucket analysis is performed to detect fraud along with operational rules on risk scores. For normal risk scores, caller profile is created with voice signature. Bad Device, Call Quality - Fraud solution uses tiered identification mechanism, which coupled with standardized voice modeling allows effective voice identification.
InsightLake call fraud solution can be deployed as an appliance in call center environment. It supports following integration points to collect call data: REST Interface with CCXML based app servers to collect call details early on CDR Logs Call Recordings Media server