Virginia Tech® home

Attacking and Securing Algorithmic Fairness in Human-Machine Interactions

Researchers will assess and improve the adversarial robustness of algorithmic fairness of artificial intelligence-powered human-machine interaction  systems, improving their value and credibility. 

Project funded by: CCI Hub

Project Investigators

Principal Investigator (PI): Jundong Li, assistant professor, University of Virginia School of Data Science, Department of Electrical and Computer Engineering, and Department of Computer Science. 

Co-PI: Jess Reia, assistant professor, University of Virginia Department of Data Science

Rationale and Background

Artificial Intelligence (AI) has empowered machines to perform and adapt to a wide spectrum of tasks, in such areas as manufacturing, health care, and autonomous driving. 

AI algorithms are often susceptible to adversarial attacks, which might entail inserting inaccurate or misrepresentative data during training, or introducing maliciously designed data to degrade its performance.

Few studies have scrutinized AI's defenses against random or malicious data attacks on algorithmic fairness.

Concerns have been raised that AI algorithms have a high risk of reproducing human biases from historical data, perpetuating discrimination against certain populations, and failing to embody inclusion, equity, and diversity.

Methodology

Researchers' have a three-phased plan: 

  • Attacking algorithmic fairness in human-machine interactions (HMI): The team will perform simulations on several benchmark object-detection algorithms in HMI.
  • Securing algorithmic fairness in HMI: The team will adapt traditional attack approaches to the setting, then evaluate the performance of the proposed securing approach with the average reduction in the change of statistical parity (or equal opportunity) before and after the attack. 
  • Establishing ethical policies for cybersecurity workforce development: The team will evaluate four aspects of the process: 
    • Participation of stakeholders in the initial stages of the project.
    • The engagement level of different stakeholders with workshops.
    • Adherence to planned dissemination of materials and resources within the stakeholders’ networks.
    • Preliminary capacity-building efforts.

Projected Outcomes

  • Improved capacity-building efforts in Virginia.
  • Advances in tackling fundamental challenges in cybersecurity, fairness AI, and human-machine interactions.