###### tags: `Reading sessions`

[toc]

---

# TIFS 2024
## [Domain Generalization via Aggregation and Separation for Audio Deepfake Detection](https://ieeexplore.ieee.org/document/10286049)
- By Yuankun Xie
- Use LCNN and Bi-LSTM extract features from voices, then creates domain classifier and train two models for detect / generate fake features. Then use Triplet mining approach to optimize the system.
- Domain Classifier: Create a feature space where the characteristics of real speech are similar, regardless of the source. It make sure the features (distinctive aspects) of real speech sound similar, even if they come from different places.
- Triplet Mining: The paper use a method called triplet mining. Imagine comparing three voice samples at a time - one real (anchor), another real (positive), and one fake (negative).
- Databases:LA Database, wavefake and fakeAVCeleb

## [Lossless Data Hiding in NTRU Cryptosystem by Polynomial Encoding and Modulation](https://ieeexplore.ieee.org/document/10423186)
- [HL] Lossless Data Hiding in Ciphertexts, LDH-CT, allows to embed data with the plain text, but not changing the plaintext, which allows to transmit extra data in the application with lower data transmission cost. The main contribution of this paper includes:
1. Propose NTRU based LDH-CT algorithms
2. A higher embedding capacity companed with NTRU (“N-th Degree Truncated Polynomial Ring Unit”)
3. Retrieve data in different scenarios using PE-PP and PM-Hybrid scheme, to improve applicability.
- [HL] The paper implements POLYNOMIAL ENCODING ALGORITHM and polynomial modulation. The PE algorithm use bit strings to generate random elements which means that it will transform a string of bit values into a polynomial r(x), and recover the extracted bit values according to the encoded polynomial. The polynomial modulation algorithm for NTRU encryption is designed to avoid making the public key invertible, as seen in the PE (Polynomial Inversion Encryption) algorithm, or reserving part of the coefficients, as in the PP (Partial Product) strategy.

## [Masked Relation Learning for DeepFake Detection](https://ieeexplore.ieee.org/document/10054130)
- [HL] Deepfake detection has following challenges:
1. Low generalizability
2. For current methods, they always includes redunant information by learning large low-informational area ( which means that they learns the entire face)
3. Current methods focusing on the inconsistency on face but is lack of finding the inconsistency in video frames
- The network implements spatiotemporal attention module and Masked Relation Learner where the video frames are divided into sequence of snippets, as every snippets has the same length D. a 3D-CNN is used to extract feature map to let the STA module produces attention maps which corresponds to specific facial region. The MRL module then discards redundant edges to focus on important relational features. It uses a Temporal Graph Convolutional Network (TGCN) to learn these features, facilitating interaction between the input graph and the hidden graph to exchange relational information. The structure Outperforms state-of-the-art approaches on detecting unseen deepfake datasets.

# 2023

[Link to accepted papers or the program]((</accepted>)https://link)

## [paper 1](https://link)
* By authors
* [AA] Short review by AA
* [BB] Short rview by BB

###### tags: `` ``

## [paper 2](https://link)
* By authors
* [AA] Short review by AA
* [BB] Short rview by BB

###### tags: `` ``
---

# [2022](https://ieeexplore.ieee.org/xpl/tocresult.jsp?isnumber=9652463&punumber=10206)

## [Black-Box Dataset Ownership Verification via Backdoor Watermarking](https://ieeexplore.ieee.org/document/10097580)
* By Yiming Li; Mingyan Zhu; Xue Yang; Yong Jiang; Tao Wei; Shu-Tao Xia
* [FH] This paper proposes to watermark datasets in such a way that one can check if the datasets have been used in DNN (deep neural network) training. The idea is similar to poison-based attacks, but the attacking methods are used for good purposes. Open source implementation: https://github.com/THUYimingLi/DVBW
* [HL] Dataset, espeically for high-quality datasets, usually should not be adopted for commercial use without permission. There is no good way of ensuring this. This research formulate a protection of released datasets, to let it determine whether the dataset is used for the model training. The database will be poisoned conditionally based on the following priciples:
* Harmless: Watermark will not be harmful to dataset functionality.
* Distinctiveness: All models thrained on watermarked dataset should have some distinctive prediction behaviours on watermarked data
* Stealthiness: Dataset watermarking should not attract attention of adversaries

The datasets this paper used includes:
1. REDDIT-MULTI-5K: A relational dataset extracted from reddit, have 5000 graph with 6 classes
2. Collab: 5000 Graphs with three possible classes

###### tags: `` ``

## [paper 2](https://link)
* By authors
* [AA] Short review by AA
* [BB] Short rview by BB

###### tags: `` ``
---

# 2021

[Link to accepted papers or the program]((</accepted>)https://link)

## [paper 1](https://link)
* By authors
* [AA] Short review by AA
* [BB] Short rview by BB

###### tags: `` ``

## [paper 2](https://link)
* By authors
* [AA] Short review by AA
* [BB] Short rview by BB

###### tags: `` ``