Abstract: Data can be received that includes information corresponding to a set of users. Privacy protection protocols that apply to the data can be identified. A subset of the data can be identified as being personally identifiable information (PII) data, where the subset includes a set of PII attributes. The PII attributes can be split into categories based on a format of a data field in the PII attributes. The processed PII data can be combined with non-PII data to create processed client data. It can be determined to add noise to part of the processed PII data. An amount of noise can be determined based on the privacy protection protocols. The amount of noise can be added to part of the processed PII data to produce protected data. A machine-learning model can be trained using the protected data.

Abstract: Techniques are disclosed for generating machine learning models that are insensitive to drift. A system trains a machine learning model using a divergent training dataset including synthesized data points simulating drift. The system can evaluate the machine learning models in terms of accuracy, latency, efficiency, and other metrics. Based on the evaluation, the system can select a machine learning model least susceptible to drift.

Abstract: A secure, modular multi-tenant machine learning platform is configured to: receive untrusted code supplied by a first tenant; perform a security scan of the untrusted code to determine whether the untrusted code satisfies a set of one or more security requirements; responsive to determining that the untrusted code satisfies the security requirement(s): deploy the untrusted code to a runtime execution environment; deploy a machine learning model associated with the first tenant to the runtime execution environment, the untrusted code being configured to perform one or more functions using the machine learning model; receive a set of untrusted code supplied by a second tenant; perform a security scan of the untrusted code to determine whether the untrusted code satisfies the security requirement(s); and responsive to determining that the untrusted code does not satisfy the security requirement(s): refraining from deploying the untrusted code to a runtime execution environment.

Abstract: For at least a selected class attribute of the multiple class attributes, one or more bias metrics are determined that estimate a degree to which a particular workflow (having a set of processing stages) is biased in association with the class attribute. Each user of a set of users is associated with a set of user data to be processed by the particular workflow. At least one of the set of processing stages includes executing a machine-learning model. It can be detected that a bias-mitigation option corresponding to a specific class attribute has been selected. For each of at least two of the set of processing stages: a de-biasing technique is selected; and the processing stage is modified by applying the de-biasing technique. A modified version of the particular workflow (which includes the modified processing stages) is applied to each of a set of input data sets.