Yifei Li

Yifei Li

Auto-generated profile, based on publicly available information

Yifei Li

University of Pennsylvania

Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
20 January 2021 in Security and Communication Networks
Verified
Journal Article
Open access
PDF file
ZCZhenwei Chen
Zhenwei Chen
AWAxin Wu
Axin Wu
YLYifei Li
Yifei Li
QXQixuan Xing
Qixuan Xing
LGLing Sheng Geng
Ling Sheng Geng
The emergence of the cloud storage has brought great convenience to people’s life. Many individuals and enterprises have delivered a large amount of data to the third-party server for storage. Thus, the privacy protection of data retrieved by the user needs to be guaranteed. Searchable encryption technology for the cloud environment is adopted to ensure that the user information is secure with retrieving data. However, most schemes only support single-keyword search and do not support file updates, which limit the flexibility of the scheme. To eliminate these problems, we propose a blockchain-enabled public key encryption scheme with multi-keyword search (BPKEMS), and our scheme supports file updates. In addition, smart contract is used to ensure the fairness of transactions between data owner and user without introducing a third party. At the data storage stage, our scheme realizes the verifiability by numbering the files, which ensures that the ciphertext received by the user is complete. In terms of security and performance, our scheme is secure against inside keyword guessing attacks (KGAs) and has better computation overhead than other related schemes.

1 January 2023
Verified
Conference Paper
Open access
PDF file
LYLi Yifei
Li Yifei
LULyle Ungar
Lyle Ungar
JSJoão Sedoc
João Sedoc
Li Yifei, Lyle Ungar, João Sedoc. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

29 September 2022
Verified
Book Chapter
Closed access
PDF file
YLYifei Li
Yifei Li
Yinghui ZhangYinghui Zhang
Yinghui Zhang
WLWei Liu
Wei Liu
Jianting NingJianting Ning
Jianting Ning
DZDong Zheng
Dong Zheng


30 October 2023 in arXiv
Verified
Preprint
Open access
PDF file
LYLi S. Yifei
Li S. Yifei
LULyle Ungar
Lyle Ungar
JSJoão Sedoc
João Sedoc
Pre-trained large language models (LLMs) reflect the inherent social biases of their training corpus. Many methods have been proposed to mitigate this issue, but they often fail to debias or they sacrifice model accuracy. We use conceptors--a soft projection method--to identify and remove the bias subspace in LLMs such as BERT and GPT. We propose two methods of applying conceptors (1) bias subspace projection by post-processing by the conceptor NOT operation; and (2) a new architecture, conceptor-intervened BERT (CI-BERT), which explicitly incorporates the conceptor projection into all layers during training. We find that conceptor post-processing achieves state-of-the-art (SoTA) debiasing results while maintaining LLMs' performance on the GLUE benchmark. Further, it is robust in various scenarios and can mitigate intersectional bias efficiently by its AND operation on the existing bias subspaces. Although CI-BERT's training takes all layers' bias into account and can beat its post-processing counterpart in bias mitigation, CI-BERT reduces the language model accuracy. We also show the importance of carefully constructing the bias subspace. The best results are obtained by removing outliers from the list of biased words, combining them (via the OR operation), and computing their embeddings using the sentences from a cleaner corpus.

1 April 2023
Verified
Conference Paper
Closed access
YLYi-Fei Li
Yi-Fei Li
SCShan Chen
Shan Chen
In short video communication, the user's emotional analysis is the basis of exploring the communication law of video media. The short video danmaku contains important information to analyze the user's emotional tendency. So identifying the emotional tendency of the audience via short video danmaku is significant for the video media communication. A novel prediction model for short video danmaku emotion recognition is designed. Firstly, the short video audience's emotions are divided into five categories to construct the user's emotional space. Then the video danmaku is preprocessed by shallow learning model. Then the preprocessed text vector generated by the preprocessor is input into Bi(LSTM+GRU)-Att model to extract high level features. Finally the shallow learning model completes the danmaku emotion recognition. The prediction model not only fully extracts the text and semantic features of the context, but also improves the model's interpretability. The experimental results show that the classification accuracy reaches 87.7% and the accuracy on the test set reaches 99.6%. Therefore this model has high accuracy in short video danmaku emotion recognition. The proposed model can be used as a basic tool for short video propagation mechanism analysis.