Projects

Project 1: DiffAdv: Generating an Adversarial Example for Any Given Image Using Diffusion Models [link]

2023.10 - 2023.12

• Motivated by the lack of efficiency and scalability of black-box adversarial attacks, we generate adversarial samples using a diffusion model.
• Our preliminary results show a drop in the accuracy of the generated adversarial examples.

Project 2: On the Role of Inductive Graph Reasoning on Improving Resilience against Backdoor Attacks to Commonsense Knowledge Graphs [link]

2022.10 - 2022.12

• We investigated the role of inductive graph reasoning on improving the resilience of Commonsense Knowledge Graph (CSKG) against backdoor attacks.
• The experiments indicated that there is a distribution shift in the prediction heads, and hence, inductive graph reasoning can enhance the model’s resilience to some extent. Nevertheless, the impact is rather modest, as the attack success rate remains virtually unchanged.

Project 3: Face Mask Recognition [link]

2022.04 - 2022.06

• We have developed an object detection dataset, with five classes of face masks.
• We achieved an MAp of 95\%+ in mask detection with YOLOv5 and Faster-RCNN.