{"id":145653,"date":"2026-07-01T08:59:39","date_gmt":"2026-07-01T00:59:39","guid":{"rendered":"https:\/\/www.curtin.edu.au\/research\/?post_type=hdr-r-projects&p=145653"},"modified":"2026-07-01T08:59:39","modified_gmt":"2026-07-01T00:59:39","slug":"synthetic-behavioural-data-for-privacy-preserving-human-state-modelling","status":"publish","type":"hdr-r-projects","link":"https:\/\/www.curtin.edu.au\/research\/hdr-r-projects\/synthetic-behavioural-data-for-privacy-preserving-human-state-modelling\/","title":{"rendered":"Synthetic Behavioural Data for Privacy-Preserving Human State Modelling"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n

Modern human-centred AI systems can infer subtle human states from behavioural and physiological signals: fatigue, stress, cognitive effort, uncertainty, impaired performance, engagement and behavioural inconsistency. These signals are valuable in safety, health, education, transport, teleoperation and training. They are also identity-rich. A pose sequence, physiological time series, gait trace, or interaction pattern may reveal who a person is, even after obvious facial or name information has been removed.<\/p>\n\n\n\n

This project will develop methods for generating synthetic behavioural data from existing human datasets while retaining useful state information and suppressing identity leakage. The data types of interest include pose, gait, skeletons, movement trajectories, wearable physiology, interaction logs and multimodal behavioural time series. The student will investigate whether generative AI can create data that is realistic enough for training and benchmarking human-state models, but private enough to reduce re-identification and membership leakage.<\/p>\n\n\n\n

The project fits Curtin’s Human-Centric AI agenda by combining responsive AI with responsible AI. Existing synthetic data tools are useful baselines, but most do not solve the harder research problem: preserving the signal needed to infer state while controlling what is leaked about identity, session, demographic attributes or dataset membership. The proposed PhD will treat synthetic behavioural data as a measurable privacy-utility problem rather than as a vague anonymisation claim.<\/p>\n\n\n\n

Aim<\/p>\n\n\n\n

    \n
  1. To substantially contribute to responsible AI by enabling useful behavioural datasets to be shared, analysed and modelled with reduced privacy risk.<\/li>\n\n\n\n
  2. To develop generative methods for creating state-conditioned synthetic behavioural data from existing real datasets.<\/li>\n\n\n\n
  3. To produce a rigorous evaluation framework that measures realism, state-inference utility and identity leakage rather than assuming that synthetic data is automatically private.<\/li>\n<\/ol>\n\n\n\n

    Objectives <\/p>\n\n\n\n

      \n
    1. Curate and characterise suitable behavioural datasets, including public wearable and physiological time-series datasets, as well as pose, gait, activity datasets, and relevant datasets already available within the research group where ethics and permissions allow.<\/li>\n\n\n\n
    2. Create baseline synthetic data generators for behavioural time series, skeleton\/pose sequences and multimodal state-labelled data.<\/li>\n\n\n\n
    3. Develop state-preserving and identity-suppressing generative models that explicitly separate the target state signal from identity and context.<\/li>\n\n\n\n
    4. Evaluate synthetic data utility by training downstream models for fatigue, effort, stress, uncertainty, engagement or related states and testing whether performance transfers to real data.<\/li>\n\n\n\n
    5. Evaluate privacy using adversarial re-identification, membership inference, attribute inference and cross-session linkage tests.<\/li>\n\n\n\n
    6. Produce guidelines, code and benchmark protocols for privacy-preserving synthetic behavioural data generation.<\/li>\n<\/ol>\n\n\n\n

      Computational techniques<\/p>\n\n\n\n