Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Steganographic Embeddings as an Effective Data Augmentation

Nicholas DiSalvo

TL;DR

The paper tackles overfitting in image classification by introducing a novel data augmentation that uses naive LSB steganography to embed a second image into each training sample. By sampling a pair of images with probability $p$ and applying $k$-bit LSB embedding, composite images are produced while maintaining the same batch structure, effectively enriching gradient information without increasing training time. The augmentation behaves as a discrete, uniformly distributed approximation of color transformations (brightness/contrast/saturation) and acts as a bit-level regularizer, promoting more robust feature learning. On CIFAR-10 with a compact CNN, steganographic augmentation yields training efficiency gains, performs competitively with standard augmentations, and provides benefits when combined with geometric transforms like horizontal flips, suggesting practical utility and potential extension to other modalities.

Abstract

Image Steganography is a cryptographic technique that embeds secret information into an image, ensuring the hidden data remains undetectable to the human eye while preserving the image's original visual integrity. Least Significant Bit (LSB) Steganography achieves this by replacing the k least significant bits of an image with the k most significant bits of a secret image, maintaining the appearance of the original image while simultaneously encoding the essential elements of the hidden data. In this work, we shift away from conventional applications of steganography in deep learning and explore its potential from a new angle. We present experimental results on CIFAR-10 showing that LSB Steganography, when used as a data augmentation strategy for downstream computer vision tasks such as image classification, can significantly improve the training efficiency of deep neural networks. It can also act as an implicit, uniformly discretized piecewise linear approximation of color augmentations such as (brightness, contrast, hue, and saturation), without introducing additional training overhead through a new joint image training regime that disregards the need for tuning sensitive augmentation hyperparameters.

Steganographic Embeddings as an Effective Data Augmentation

TL;DR

The paper tackles overfitting in image classification by introducing a novel data augmentation that uses naive LSB steganography to embed a second image into each training sample. By sampling a pair of images with probability and applying -bit LSB embedding, composite images are produced while maintaining the same batch structure, effectively enriching gradient information without increasing training time. The augmentation behaves as a discrete, uniformly distributed approximation of color transformations (brightness/contrast/saturation) and acts as a bit-level regularizer, promoting more robust feature learning. On CIFAR-10 with a compact CNN, steganographic augmentation yields training efficiency gains, performs competitively with standard augmentations, and provides benefits when combined with geometric transforms like horizontal flips, suggesting practical utility and potential extension to other modalities.

Abstract

Image Steganography is a cryptographic technique that embeds secret information into an image, ensuring the hidden data remains undetectable to the human eye while preserving the image's original visual integrity. Least Significant Bit (LSB) Steganography achieves this by replacing the k least significant bits of an image with the k most significant bits of a secret image, maintaining the appearance of the original image while simultaneously encoding the essential elements of the hidden data. In this work, we shift away from conventional applications of steganography in deep learning and explore its potential from a new angle. We present experimental results on CIFAR-10 showing that LSB Steganography, when used as a data augmentation strategy for downstream computer vision tasks such as image classification, can significantly improve the training efficiency of deep neural networks. It can also act as an implicit, uniformly discretized piecewise linear approximation of color augmentations such as (brightness, contrast, hue, and saturation), without introducing additional training overhead through a new joint image training regime that disregards the need for tuning sensitive augmentation hyperparameters.

Paper Structure

This paper contains 13 sections, 8 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Related Works
  3. Steganography as an Augmentation
  4. The Algorithm
  5. Efficiency of Steganography as an Augmentation
  6. Steganography as a Discrete Linear Approximator of Color Augmentations
  7. Additional Benefits of Steganographic Augmentations
  8. Experimental Details
  9. Dataset, Loss Function, and Model Architectures
  10. Optimization
  11. Steganography Hyperparameters
  12. Results
  13. Conclusion

Figures (4)

  • Figure 1: This is a visual demonstration of the LSB Steganography Augmentation Strategy. When creating batches for our dataset, we at random take image pairs from the dataset and one inside the other to create an embedded steganographic image we will use to train our model.
  • Figure 2: Number of Bits For Steganography: [1,2,3,5,7] From this visual, we see stegangraphic embeddings implicitly add color augmentations such as brightness, contrast, saturation, and hue to our model with no added overhead.
  • Figure 3: Results comparing steganographic augmentations to color-jitter (saturation, brightness, contrast, hue), Gaussian blur, random erasing, horizontal flips, no augmentations, and a combined horizontal flip + steganography
  • Figure :