ABSTRACT

Generative models describe real-world data distributions such as images, texts, and human motions, and are playing an essential role in a large and growing range of applications from photo editing to natural language processing to autonomous driving. There are two open challenges regarding the development and dissemination of generative models: (1) Adversarial applications of generative models have created concerning socio-technical disturbances (e.g., espionage operations and malicious impersonation); and (2) developing generative models using multiple proprietary datasets (which are needed to reduce data biases) raises privacy concerns about data leakage. Legislative efforts have recently been taken in the wake of these challenges, so far with limited consensus on the format of regulations and knowledge about their technological or social feasibility. To this end, this project will develop new mathematical theories and computational tools to assess the feasibility of two connected solutions to these challenges: Model attribution enforces the owners to be correctly identified based on their generated contents; secure training ensures zero data leakage during the collaborative training of attributable generative models. If successful, the outcomes of the project will provide technical guidance for future regulation design towards secure development and dissemination of generative models. Project results will be disseminated through a project website, open-source software, and public datasets. The impacts of the project will be broadened through educational activities, including new course modules on Artificial Intelligence (AI) security, undergraduate research projects, and outreach to the local community through lab tours, to prepare underrepresented groups with skills to mitigate risks from malicious impersonation and biased data/model representations targeting these groups.

This project will focus on synergistic research tasks towards decentralized model attribution and secure training of generative models. In the former, the research team will study the systematic design of a set of user-end generative models that can be certifiably attributed by a set of binary classifiers, which are stored in a decentralized manner to mitigate security risks. The technical feasibility of decentralized attribution will be measured by the tradeoffs between attributability, generation quality, and model capacity. In the latter, the research team will study secure multi-party training of generative models and the associated binary classifiers for attribution. Data privacy and training scalability will be balanced through the design of security-friendly model architectures and learning losses. New knowledge will be created that differentiates this project from the existing state-of-the-art literature in digital forensics and secure computation: (1) Sufficient conditions for decentralized attribution will be developed, which will reveal analytical connections between attributability, data geometry, model architecture, and generation quality. (2) The sufficient conditions will enable estimation of the capacity of attributable models for a given dataset and generation quality tolerance. (3) Feasibility of sublinear secure vector multiplication will be studied, which will fundamentally improve the scalability of secure collaborative training. (4) Privacy-friendly activation and loss functions will be designed for the training of user-end generative models and the classifiers for attribution.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Please report errors in award information by writing to: awardsearch@nsf.gov.