Our approach generates images based on the given text-prompts and camera settings -- numerical values embedded as tokens in the aligned text feature space of corresponding Latent Diffusion Models (LDMs). This approach enables more effective and precise control on image generation.

We explore how to enhance the generalization of deepfake face video recognition models. Building on the results of my master's thesis, we propose a dual-branch side network architecture adapted to the CLIP image foundation model, combined with a facial component attention loss function. This approach successfully improves the model and achieves promising generalization performance.

We propose to enhance the generalizability of the self-supervised deepfake detector through multi-task learning. This extension on self-supervised learning enables the detector to utilize reasonable information for authenticity evaluation on the given image. This simple method demonstrates effectiveness through substantial improvements in the evaluation on unseen deepfake datasets.

We study the generalization issues of detection models for spoofing and deepfake facial images. We propose a meta-learning methods to initialize the weights of a lightweight classifier to enhance its generalizability.

We extend previous results on in social vehicular networks by exploring in depth the impact of the size and number of social communities to the quality of the wireless vehicular network. Further, we also propose a tractable greedy resource allocation algorithm to tackle the excessive computation time of the optimal algorithm for the multi-group multi-QoS Non-Orthogonal Multi-Access (NOMA) resource allocation problem.

This work is the beginning of my research journey. I followed my advisor to research on resolving the impact of bursty traffic in traffic conjections under 5G vehicular wireless communication context. I studied about wireless communication and refer to the 3GPP standard to design a simulation model for wireless communication. Further, to resolve problem of optimal resource allocation, I disentangle the problem in order to solve it through utilizing the integer linear programming technique. The simulation system incorporates SUMO, Matlab with python as the bridge to control the two components.