Bi-level Optimization-based Pruning

Supervisor: Dr. Sijia Liu (MSU) Collabrator: Dr. Parikshit Ram (IBM), Dr. Mingyi Hong (UMN)

  • Devise the model pruning algorithm through the lens of bi-level optimization.

  • Colead and build the new bi-level pruning (BIP) pipeline, achieving SOTA performance.

Reverse Engineering of Adversarial Deceptions

Supervisor: Dr. Sijia Liu (MSU) Collabrator: Dr. Xiaoming Liu (MSU), Dr. Xue Lin (NEU)

  • Design Reverse Engineering of Deceptions (RED) pipeline to recover adversarial perturbations and infer victim model attributes.

  • Lead and build the first RED framework, achieving better performance than denoising methods.

Scalable Optimization for Adversarial Learning.

Supervisor: Dr. Sijia Liu (MSU) Collabrator: Dr. Shiyu Chang (UCSB), Dr. Pin-Yu Chen (IBM)

  • Build the black-box defense as a query-based optimization system to ensure certified robustness.

  • Integrating the bi-level optimization in to the backdoor example generation process.

  • Build a MAML-based few-shot universal adversarial perturbation generation method.