Authenticating users in the presence of small transaction volumes
A method involves performing a mathematical estimation operation identifying a risk score threshold. The operation identifies the risk score threshold as a point on a curve rather than a value of a particular risk score. Such a curve approximates the distribution of risk score values output over a time interval and represents a function embodied by a plot of risk score percentile vs. risk score value. The risk engine, rather than selecting a particular risk score, selects a curve from a family of curves that is known to accurately represent such risk score distributions. For example, the risk engine may choose the curve that provides the best fit to the previous week's risk scores over the family of curves. The risk engine identifies the risk score threshold by finding a risk score value such that the function evaluated at that risk score value produces a specified risk score percentile.
1. A method of selectively providing access to users who have correctly identified themselves, the method comprising:
(a) during a first time interval:
receiving, by a resource server, a plurality of resource access requests from user devices over a network,
in response to each of the received plurality of resource access requests, generating a respective authentication request based on values of authentication factors from that resource access request and sending that authentication request from the resource server to an authentication server over the network,
generating, by processing circuitry of the authentication server, a plurality of risk scores in response to receiving the authentication requests;
(b) deriving, by the processing circuitry, a risk score threshold from the risk scores by:
generating a fitted parametric distribution function of risk score percentile versus risk score by fitting the risk scores to a model of a parametric distribution function,
identifying a point along the fitted parametric distribution function having a first coordinate equal to a threshold risk score percentile and a second coordinate that is not found among the risk scores generated by the processing circuitry during the first time interval, and
determining the risk score threshold to be equal to the second coordinate; and
(c) during a second time interval subsequent to the first time interval:
receiving, by the resource server from a first user device over the network, a first resource access request to access a first protected account hosted by the resource server,
in response to the resource server receiving the first resource access request, generating a first authentication request based on the first resource access request and sending the first authentication request from the resource server to the authentication server over the network, the first authentication request containing first values of authentication factors,
receiving, by the resource server from a second user device over the network, a second resource access request to access a second protected account hosted by the resource server,
in response to the resource server receiving the second resource access request, generating a second authentication request based on the second resource access request and sending the second authentication request from the resource server to the authentication server over the network, the second authentication request containing second values of authentication factors,
generating, by the processing circuitry, a first risk score based on the first values of the authentication factors of the first authentication request and a second risk score based on the second values of the authentication factors of the second authentication request;
determining, by the processing circuitry, that the first risk score is less than the derived risk score threshold, and, in response, providing the first user device with access to the first protected account; and
determining, by the processing circuitry, that the second risk score is greater than the derived risk score threshold, and, in response, invoking an action prior to providing or denying the access to the second protected account to the second user device, wherein invoking the action includes:
requesting additional authentication factors from the second user device; and
performing follow-up authentication based on the additional authentication factors prior to determining whether the second user device may access the second resource.
2. A method as in claim 1 , wherein the model of the parametric distribution function includes a variable set of parameters; and
wherein fitting the risk scores to the model includes finding a particular value of each of the set of parameters based on the risk scores.
3. A method as in claim 2 , wherein fitting the risk scores to the model further includes rescaling the risk scores to respective nonnegative rescaled risk scores; and
wherein finding the particular value of each of the set of parameters using the risk scores includes performing a maximum likelihood estimation of the set of parameters using the rescaled risk scores.
4. A method as in claim 3 , wherein rescaling the risk scores includes:
generating an interquartile range of the risk scores; and
dividing each of the risk scores by the interquartile range in order to avoid scaling the risk scores by outliers.
5. A method as in claim 3 , wherein the fitted parametric distribution function is equal to a first distribution function of rescaled risk score when a rescaled risk score is less than a value of a cutoff parameter of the set of parameters and a second distribution function of rescaled risk score when the rescaled risk score is greater than the value of the cutoff parameter, the second distribution function representing a tail of a distribution of the rescaled risk scores; and
wherein performing the maximum likelihood estimation of the set of parameters using the rescaled risk scores includes generating a log-likelihood function of the set of parameters from the first distribution function, the second distribution function, and the rescaled risk scores.
6. A method as in claim 5 , wherein the first distribution function of rescaled risk score is a gamma distribution function, and the second distribution function of rescaled risk score is proportional to a generalized Pareto distribution function;
wherein the gamma distribution function has a gamma shape parameter, k, and a gamma scale parameter, θ;
wherein the generalized Pareto distribution function has a Pareto shape parameter, ξ, a Pareto scale parameter, σ u , and a location parameter, u;
wherein the method further comprises setting the cutoff parameter to be equal to the location parameter, u; and
wherein generating the log-likelihood function includes normalizing the generalized Pareto distribution function to cause the distribution function of risk score to be equal to one for risk scores sufficiently greater than the risk score threshold.
7. A system constructed and arranged to selectively provide access to users who have correctly identified themselves, the system comprising:
a resource server including first controlling circuitry coupled to first memory; and
an authentication server including second controlling circuitry coupled to second memory;
wherein the first memory stores first instructions, which, when executed by the first controlling circuitry, cause the resource server to, during a first time interval:
receive a plurality of resource access requests from user devices over a network and
in response to each of the received plurality of resource access requests, generate a respective authentication request based on values of authentication factors from that resource access request and send that authentication request to the authentication server over the network;
wherein the second memory stores second instructions, which, when executed by the second controlling circuitry cause the authentication server to:
during the first time interval, generate a plurality of risk scores in response to receiving the authentication requests and
derive a risk score threshold from the risk scores by:
generating a fitted parametric distribution function of risk score percentile versus risk score by fitting the risk scores to a model of a parametric distribution function,
identifying a point along the fitted parametric distribution function having a first coordinate equal to a threshold risk score percentile and a second coordinate that is not found among the risk scores generated by the authentication server during the first time interval, and
determining the risk score threshold to be equal to the second coordinate;
wherein the first instructions, when executed by the first controlling circuitry, further cause the resource server to, during a second time interval subsequent to the first time interval:
receive from a first user device over the network, a first resource access request to access a first protected account hosted by the resource server,
in response to the resource server receiving the first resource access request, generate a first authentication request based on the first resource access request and send the first authentication request to the authentication server over the network, the first authentication request containing first values of authentication factors,
receive from a second user device over the network, a second resource access request to access a second protected account hosted by the resource server, and
in response to the resource server receiving the second resource access request, generate a second authentication request based on the second resource access request and send the second authentication request from the resource server to the authentication server over the network, the second authentication request containing second values of authentication factors; and
wherein the second instructions, when executed by the second controlling circuitry, further cause the authentication server to, during the second time interval subsequent to the first time interval:
generate a first risk score based on the first values of the authentication factors of the first authentication request and a second risk score based on the second values of the authentication factors of the second authentication request;
determine that the first risk score is less than the derived risk score threshold, and, in response, provide the first user device with access to the first protected account; and
determine that the second risk score is greater than the derived risk score threshold, and, in response, invoke an action prior to providing or denying the access to the second protected account to the second user device, wherein invoking the action includes:
requesting additional authentication factors from the second user device; and
performing follow-up authentication based on the additional authentication factors prior to determining whether the second user device may access the second resource.
8. A system as in claim 7 , wherein the model of the parametric distribution function includes a variable set of parameters; and
wherein the second instructions, when executed by the second controlling circuitry, when fitting the risk scores to the model, further cause the authentication server to find a particular value of each of the set of parameters based on the risk scores.
9. A system as in claim 8 , wherein the second instructions, when executed by the second controlling circuitry, when fitting the risk scores to the model, further cause the authentication server to rescale the risk scores to respective nonnegative rescaled risk scores; and
wherein the second instructions, when executed by the second controlling circuitry, when finding the particular value of each of the set of parameters using the risk scores, further cause the authentication server to perform a maximum likelihood estimation of the set of parameters using the rescaled risk scores.
10. A system as in claim 9 , wherein the second instructions, when executed by the second controlling circuitry, when rescaling the risk scores, further cause the authentication server to:
generate an interquartile range of the risk scores; and
divide each of the risk scores by the interquartile range in order to avoid scaling the risk scores by outliers.
11. A system as in claim 9 , wherein the fitted parametric distribution function is equal to a first distribution function of rescaled risk score when a rescaled risk score is less than a value of a cutoff parameter of the set of parameters and a second distribution function of rescaled risk score when the rescaled risk score is greater than the value of the cutoff parameter, the second distribution function representing a tail of a distribution of the rescaled risk scores; and
wherein the second instructions, when executed by the second controlling circuitry, when performing the maximum likelihood estimation of the set of parameters using the rescaled risk scores, further cause the authentication server to generate a log-likelihood function of the set of parameters from the first distribution function, the second distribution function, and the rescaled risk scores.
12. A system as in claim 11 , wherein the first distribution function of rescaled risk score is a gamma distribution function, and the second distribution function of rescaled risk score is proportional to a generalized Pareto distribution function;
wherein the gamma distribution function has a gamma shape parameter, k, and a gamma scale parameter, θ;
wherein the generalized Pareto distribution function has a Pareto shape parameter, ξ, a Pareto scale parameter, σ u , and a location parameter, u;
wherein the second instructions, when executed by the second controlling circuitry, further cause the authentication server to set the cutoff parameter to be equal to the location parameter, u; and
wherein the second instructions, when executed by the second controlling circuitry, when generating the log-likelihood function, further cause the authentication server to normalize the generalized Pareto distribution function to cause the distribution function of risk score to be equal to one for risk scores sufficiently greater than the risk score threshold.