Perfectly Secure Steganography Using Minimum Entropy Coupling
Authors: Christian Schroeder de Witt, Samuel Sokota, J Zico Kolter, Jakob Nicolaus Foerster, Martin Strohmeier
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To provide empirical validation, we compare a minimum entropy coupling-based approach to three modern baselines arithmetic coding, Meteor, and adaptive dynamic grouping using GPT-2, Wave RNN, and Image Transformer as communication channels. We find that the minimum entropy coupling-based approach achieves superior encoding efficiency, despite its stronger security constraints. |
| Researcher Affiliation | Collaboration | Christian Schroeder de Witt FLAIR, University of Oxford cs@robots.ox.ac.uk Samuel Sokota Carnegie Mellon University ssokota@andrew.cmu.edu J. Zico Kolter Carnegie Mellon University zkolter@cs.cmu.edu Jakob Foerster FLAIR, University of Oxford jakob@robots.ox.ac.uk Martin Strohmeier armasuisse Science + Technology martin.strohmeier@armasuisse.ch |
| Pseudocode | Yes | Algorithm 1 Sampling S γ(C | X=x) Algorithm 2 Computing γ(X | C=s) |
| Open Source Code | Yes | An open source implementation of i MEC is available at https://github.com/ schroederdewitt/perfectly-secure-steganography. |
| Open Datasets | Yes | The second and third covertext distributions are variants of GPT-2 (Radford et al., 2019) with 12 attention modules (Wolf et al., 2020) conditioned on 1024-character strings from the Wikitext-103 dataset (Merity et al., 2016). The fourth covertext distribution consists of a text-to-speech (TTS) pipeline (Yang et al., 2022) based on Tacotron-2 (Shen et al., 2018) mel-spectogram encodings, followed by a Wave RNN generative model (Kalchbrenner et al., 2018) pretrained on the LJSpeech dataset (Ito & Johnson, 2017) with a mean covertext entropy of HC = 0.35 bits per token. The fifth covertext distribution is over an unconditional transformerbased generative model (Parmar et al., 2018) of CIFAR-10 (Krizhevsky et al., 2009) RGB images of dimensions 32 32. |
| Dataset Splits | No | The paper does not provide specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) for training, validation, or testing for their experiments. It refers to using pre-trained models and existing datasets as covertext distributions for evaluation, rather than defining new splits for training their own models from scratch. |
| Hardware Specification | Yes | All experiments were performed on a AMD Ryzen Threadripper PRO 3955WX with 16 physical cores and 2x NVIDIA Ge Force RTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions several models and frameworks used (e.g., GPT-2, Wave RNN, Tacotron-2, Torch Audio) and cites papers for them, but it does not provide specific version numbers for these software components or other general dependencies (like Python, PyTorch, etc.) that would be needed for reproducible setup. |
| Experiment Setup | Yes | Our implementation of i MEC makes use of the approximative minimum entropy coupling procedure described in Algorithm 1 of (Kocaoglu et al., 2017) as its subprocedure. We show results for three different i MEC block sizes (10 bits, 16 bits, and 20 bits). The number of bits measures the initial entropy of a particular H(µi). We stop transmission when the entropy of every posterior block is less than 0.1, i.e., i, H(µi) < 0.1. For arithmetic coding and Meteor/Meteor:reorder, we performed a grid search over precision values of 8, 16, 32 and 48 bits. All ciphertexts are 80-bit bitstrings sampled uniformly at random. We tuned the hyper-parameters of each method to yield the best performance on this task. |