3-party adversarial steganography

Ishak Meraouche, Sabyasachi Dutta, Kouichi Sakurai

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

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.

Original languageEnglish
Title of host publicationInformation Security Applications - 21st International Conference, WISA 2020, Revised Selected Papers
EditorsIlsun You
PublisherSpringer Science and Business Media Deutschland GmbH
Pages89-100
Number of pages12
ISBN (Print)9783030652982
DOIs
Publication statusPublished - 2020
Event21st International Conference on Information Security Applications, WISA 2020 - Jeju Island, Korea, Republic of
Duration: Aug 26 2020Aug 28 2020

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12583 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference21st International Conference on Information Security Applications, WISA 2020
CountryKorea, Republic of
CityJeju Island
Period8/26/208/28/20

All Science Journal Classification (ASJC) codes

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint Dive into the research topics of '3-party adversarial steganography'. Together they form a unique fingerprint.

Cite this