Rethinking Data Bias: Dataset Copyright Protection via Embedding Class-wise Hidden Bias

Rethinking Data Bias: Dataset Copyright Protection via Embedding Class-wise Hidden Bias

논문 정보

• Jinhyeok Jang, BungOk Han, Jaehong Kim, Chan-Hyun Youn
• Computer Vision - ECCV 2024
• ETRI (Electronics and Telecommunications Research Institute) / KAIST (Korea Advanced Institute of Science and Technology)

1. Introduction

2. Related Work

2-1. Backdoor Attacks

2-2. Data Poisoning

2-3. Radioactive Data

2-4. Limitations of the Prior Works in Verification

3. Motivation

4. Method

4-1. Noise Patch Placement: Class-wise Bias Embedding

4-2. Overlaying Auxiliary Dataset: Robust Bias to Augmentation

4-3. Undercover Bias: Invisible Bias Embedding

4-4. Discussion

5. Experiments I: Comparison with Prior Works

5-1. Comparison in Fundamental Specifications

5-2. Comparison in Effectiveness of Watermark

6. Experiments II: General Applicability

6-1. Application to Further Architectures and Datasets

6-2. Application to Fine-grained Classification

6-3. Application to Image Segmentation

7. Ablation Studies

7-1. Histogram Analysis of mAcc on Watermark

7-2. Visualizations

8. Conclusion