What Time Tells Us? Learning time-Awareness from Static Images, Dongheng Lin^*, Han Hu^*, Jianbo Jiao^†

• A new paradigm to learn visual cues for timestamps leading to time-aware understanding on random visual inputs.
• Built a new benchmark dataset with 130k accurate samples paired with metadata.
• Tested the learned representations on various downstream tasks (Image Retrieval, Video Scene Recognition and Time-conditioned Image Editing).

SPSS: A Salience-based Poisoning Selection Strategy for Selecting Backdoor Attack Victims, Zihan Lyu¹, Dongheng Lin¹, Jie Zhang^†, Ruiming Zhang²

• Designed an algorithm uses Salience Metric to evaluate sample feature significance towards backdoor learning process.
• With its assistance, we managed to realize a more data-efficient backdoor attack to DNN models achieved the same attack success rate to vanilla backdoor attack with only 38.44% of poisoned samples.

Conditional Metadata Embedding Data Preprocessing Method for Semantic Segmentation

Juntuo Wang¹, Qiaochu Zhao¹, Dongheng Lin¹, Erick Purwanto^†, Ka Lok Man²

• In this paper, we propose a conditional data preprocessing strategy, i.e., Conditional Metadata Embedding (CME) data preprocessing strategy. The CME data preprocessing method will embed conditional information to the training data, which can assist the model to better overcome the differences in the datasets and extract useful feature information in the images.

Arxiv Explorer: An efficient paper recommendation system we developed on a full dataset of all available Arxiv Papers (until Feb 2024).

• All papers are compressed into paper embeddings using category, abstract, and title, stored in a FAISS indexing system.
• We modeled a coauthorship map to conduct authority/hub based HITS reranking of the paper recommendation results.

MapleJuice: A light-weight counterpart of Hadoop supported with SQL-like query

• The distributed system is built upon a self-implemented file system similar to GFS with corresponding NameNode and DataNode.
• We also implemented an efficient Gossip-style failure detection protocol to maintain all the node status using UDP packets, a Bully-algorithm based re-election ensures new leader will be available in case of any failures on master nodes.
• The task scheduling mechanism is similar to MapReduce, ensures the parallelism among nodes. We tested the system against Hadoop within a cluster with 10 VMs. The MapleJuice is generally 25% faster than Hadoop when dealing with small clusters.

My undergraduate thesis: StyleDiffuser: Cartoon-Style Image Creation with Diffusion Model and GAN Fusion

• In this work, I have introduced a novel approach that fuses Generative Adversarial Networks (GANs) with the Stable Diffusion Model for creating cartoon-style images.
• Utilizes StyleGAN2 generated feature maps and corresponding metadata to constrain the Stable Diffusion Network.
• Reduces the number of diffusion steps required for the model to converge to a final image, streamlining the image generation process.
• The method reduces the reliance on verbose prompts for controlling the output, making the generation process more straightforward.

• FUJICHROME Velvia 100, taken at my hometown.