2022 Fiscal Year Final Research Report
Multimedia Forensics for Classification of Fake Content
| Project/Area Number |
19K22846
|
|---|
| Research Category |
Grant-in-Aid for Challenging Research (Exploratory)
|
| Allocation Type | Multi-year Fund |
|---|
| Review Section |
Medium-sized Section 60:Information science, computer engineering, and related fields
|
|---|
| Research Institution | Okayama University |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2019-06-28 – 2023-03-31
|
| Keywords | フェイクコンテンツ / 生成AI / 画像識別器 / 敵対的サンプル |
|---|
| Outline of Final Research Achievements |
We designed and evaluated a basic image classifier for fake content created by replacing a person's face with another person's face or changing facial expressions, as well as for images and videos artificially created by an adversarial generative network. We also proposed extensions to the region-based processing and color space correction methods to achieve high classification accuracy for content created with different methods during training and testing. In addition, we developed a method to identify the presence or absence of adversarial noise generated to misidentify the image classifier. We designed a preprocessing filter that does not modify the original image classifier and confirmed through experiments that it can detect adversarial images with a high accuracy of over 90%.
|
| Free Research Field |
マルチメディアセキュリティ
|
| Academic Significance and Societal Importance of the Research Achievements |
深層学習技術を用いれば，偽物とは簡単に判別できない画像や映像を生成することが可能となっており，悪用されることが懸念されている．特に，特定の人物を操った形でのコンテンツが精巧に作成され，社会的な影響を与えるような印象操作に使われる可能性も高まっている．本研究で開発した識別器は，そのような人工的に作成および加工編集されたコンテンツを機械的に判別するものであり，マルチメディアコンテンツに関連したセキュリティ技術として重要な役割を担うものと考えられる．本識別器を拡張させて更に精度を高める研究が広がることで，フェイク情報の流布を抑制できることが期待される．
|
