A New Frontier in AI: Training AI Applications Using Sensitive Information and the Potential Data Privacy Risks Involved (RECORDING)

Recently, companies have built a new frontier of AI applications that perform predictive tasks using sensitive information. For example, today there are AI applications that predict a patient’s health issues or infer the next sentence to write in a private email. Building AI applications by training AI models with sensitive information, however, presents significant data privacy risks. If improperly trained, an AI model can unintentionally “memorize” rare or unique sensitive information included in a large set of training data. This creates the risk that the AI model will disclose the sensitive information “as-is” in response to new, previously unseen input data. This is particularly problematic for AI applications that are widely used by the public because hackers can exploit this vulnerability to extract the sensitive information. Hackers can input different prefixed texts to “inferentially attack” an improperly trained AI application to extract the sensitive information that was included in the training data set during the training process. This unintended “memorization” is also a threat to AI models trained to infer word associations from a business’s internal information. A simple inference of uncommon word combinations from the business’s internal information can reveal trade secrets. 

This webinar will describe the issues involved with training AI models using sensitive information. Attendees will also learn about the data privacy risks and legal implications that arise when AI models are trained using sensitive information. Lastly, this webinar will discuss data science techniques that can prevent the problem of “unintended memorization.”