3-party adversarial steganography

Ishak Meraouche, Sabyasachi Dutta, Kouichi Sakurai

研究成果: 書籍/レポート タイプへの寄稿会議への寄与

2 被引用数 (Scopus)


Steganography enables a user to hide information by embedding secret messages within other non-secret texts or pictures. Recently, research along this direction has picked a new momentum when Hayes & Danezis (NIPS 2017) used adversarial learning to generate steganographic images. In adversarial learning, two neural networks are trained to learn to communicate securely in the presence of eavesdroppers (a third neural network). Hayes–Danezis forwarded this idea to steganography where two neural networks (Bob & Charlie) learn “embed” and “extract” algorithms by exchanging images with hidden text in presence of an eavesdropping neural network (Eve). Due to non-convexity of the models in the training scheme, two different machines may not learn the same embedding and extraction model even if they train on the same set of images. We take a different approach to address this issue of “robustness” in the “decryption” process. In this paper, we introduce a third neural network (Alice) who initiates the process of learning with two neural networks (Bob & Charlie). We implement and demonstrate through experiments that it is possible for Bob & Charlie to learn the same embedding and extraction model by using a new loss function and training process.

ホスト出版物のタイトルInformation Security Applications - 21st International Conference, WISA 2020, Revised Selected Papers
編集者Ilsun You
出版社Springer Science and Business Media Deutschland GmbH
出版ステータス出版済み - 2020
イベント21st International Conference on Information Security Applications, WISA 2020 - Jeju Island, 韓国
継続期間: 8月 26 20208月 28 2020


名前Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
12583 LNCS


会議21st International Conference on Information Security Applications, WISA 2020
CityJeju Island

!!!All Science Journal Classification (ASJC) codes

  • 理論的コンピュータサイエンス
  • コンピュータサイエンス一般


「3-party adversarial steganography」の研究トピックを掘り下げます。これらがまとまってユニークなフィンガープリントを構成します。