Abstract

Children with autism spectrum disorders (ASD) benefit from specialized services throughout their lifespan. However, autism assessment, treatment, and care are bottlenecked by clinician availability. While it is important to continue to seek to expand the ASD-healthcare workforce, recent efforts have begun to assess alternative approaches such as crowdsourcing and computational methods for their potential to accelerate and increase accessibility to the diagnostic and therapeutic process. The development and evaluation of these methods requires access to rich data on children with ASD, specifically, audiovisual recordings, i.e., videos. However, videos contain identifiable information: the face and voice of the child. Our long-term goal is to address the technical challenge of protecting facial and vocal identity for child subjects while enabling access to rich data to improve our understanding of ASD and expand availability of treatment and care. Our current objective is to elucidate the privacy-utility tradeoffs for three approaches to privatization in the use case context of behavioral assessment of children. The premise of this study is that recent strides in generative models and adversarial machine learning have yielded deep neural network architectures that can modify faces and voices to change identity; however, these architectures have been developed and tested for contexts that involve typically developing adults. We hypothesize that these architectures can be adapted to privatize video recordings of children with ASD while retaining information essential to the quantification of ASD-associated behaviors. The significance of this study comes from the value of adapting the neural network models appropriately and specifically to the autism context as this requires the preservation of attributes that would not have an equally high importance in other use case contexts. In this proposal, we will (1) generate a rich multi- camera, multi-microphone dataset of children (5-12 years old) undergoing behavioral assessment for autism, (2) determine the impact of adaptations on the efficacy of anonymization relative to baseline implementations for three classes of anonymization methods, (3) evaluate how well ASD-associated behaviors are preserved during the privatization process for the purpose of upcoming approaches to aid the diagnostic process, specifically, crowdsourced assessments and computational assessments, and (4) conduct user centered co- design to surface guidelines on making a privatization toolkit easy to use for parents and clinicians. The outcome of this study will be a framework for evaluating facial and vocal anonymization in the context of children with ASD, development of a rich and annotated shareable video research dataset for computational or behavioral investigation, evaluation of specific methods that can make rich, diverse, heterogenous data more easily available to the autism community, and development of practical, usable workflows for anonymization for care providers. More generally, this study will pave the way for worry-free video data sharing for mental health.

We acknowledge funding from: National Institute for Mental Health “AVAIL:  Anonymization of Videos using AI for Large scale data sharing” (2024-2029).

Publications

In progress