SparSamp Artifact Description

This repository contains the Artifact for the paper "SparSamp: Efficient Provably Secure Steganography Based on Sparse Sampling".

SparSamp introduces a novel steganography scheme leveraging sparse sampling to achieve efficient and provably secure information hiding.

Overview

The Artifact provides a Python implementation for encoding and decoding messages using SparSamp within neural generative models. Core functionalities:

• Encoding: Embed messages into generated tokens via encode_spar.

• Decoding: Extract messages from tokens via decode_spar.

Used Models & Datasets

• Models:

  • Text: GPT-2 (openai-community/gpt2), Qwen-2.5 (Qwen/Qwen2.5-3B-Instruct), Llama-3 (meta-llama/Llama-3.1-8B-Instruct)

  • Image: DDPM (FFHQ dataset)

  • Audio: WaveRNN

• Datasets: IMDB text samples (first 3 sentences per sample).

Key Features

1. Provable Security

   • Preserves original probability distributions (KLD = 0).

2. High Efficiency

   • (O(1)) time complexity per sampling step.

   • Embedding speed up to 755 bits/s (GPT-2).

3. Practicality

   • Plug-and-play design: Replace sampling components in existing models.

   • Supports multi-modal tasks (text, image, audio).

Requirements

Hardware

• CPU: Intel Xeon Gold 6330 @ 2.00GHz (minimum)

• GPU: NVIDIA RTX 4090 (recommended for acceleration)

• Memory: ≥128GB RAM

• Disk: ≥20GB for model checkpoints.

Software

• Python: 3.8.5

Libraries:

torch==2.2.2
transformers==4.41.2
scipy>=1.10

Note the problems that hardware may cause.