Leaking Sensitive Financial Accounting Data in Plain Sight using Deep Autoencoder Neural Networks

Marco Schreyer, Christian Schulze, Damian BortharXiv 2020 · arXiv

Introduction: The Invisible Insider Threat

In modern corporate environments, financial data is often considered the most critical asset. Organizations invest heavily in Enterprise Resource Planning (ERP) systems and Computer-Assisted Audit Techniques (CAATs) to monitor transactions and prevent fraud. However, a significant vulnerability remains: the insider threat. An employee with legitimate access can extract sensitive tabular data—such as vendor lists or strategic pricing—and smuggle it out of the organization.

This paper introduces a novel threat model where deep steganography is used to hide massive amounts of ERP data inside ordinary, "day-to-day" images (e.g., photos of animals or food). Because these "container" images are visually indistinguishable from the originals, they easily bypass traditional Data Leakage Prevention (DLP) systems and internal audits.

Figure 1: Overview of the proposed threat model, illustrating how sensitive ERP data is hidden within images to bypass organizational boundaries and auditors.

Deep Steganography Architecture

The researchers adapted a deep neural network architecture designed to hide a full-color image (or in this case, formatted binary data) within another image of the same size. The system consists of three distinct components trained end-to-end:

  1. Preparation Network ($P_\theta$): This network transforms the "secret" data—pre-processed from tabular rows into binary bitmask images—into features that are more easily embedded within a cover image.
  2. Hiding Network ($H_\phi$): This component receives both the prepared secret features and a "cover" image. It generates a "container" image that looks identical to the cover but contains the hidden payload.
  3. Reveal Network ($R_\gamma$): Operating outside the organizational boundaries, this decoder takes the container image and reconstructs the original secret binary data.
Figure 2: The deep steganography pipeline, showing the Preparation, Hiding, and Reveal networks used to encode and decode sensitive journal entry data.

Experimental Results

To validate the approach, the authors used payment datasets from the cities of Philadelphia and Chicago as secret data, and Tiny ImageNet for cover images. The results demonstrate a highly effective and stealthy attack:

  • High Fidelity: The Structural Similarity Index Measure (SSIM) was consistently 0.99 or higher, meaning the container images are virtually identical to the originals. The Peak Signal-to-Noise Ratio (PSNR) reached 42-43 dB, a level where human observers cannot detect any tampering.
  • Near-Perfect Recovery: The Reveal network achieved a Bit Accuracy (BACC) of over 99.7%, ensuring that the leaked financial records remain actionable and accurate after extraction.
  • Learned Obfuscation: Unlike traditional steganography, which modifies specific bits in a predictable pattern, this deep learning approach spreads information across all color channels and pixel intensities, making it statistically difficult for standard tools to detect.

Conclusion and Future Work

This research highlights a critical gap in current internal audit practices. While auditors are trained to look for suspicious transactions, they are rarely equipped to scan non-financial files for hidden data.

The authors suggest that future work must focus on steganalysis—developing machine learning models specifically designed to detect the subtle statistical signatures left by deep autoencoders. Integrating these detection capabilities into internal audit and fraud examination workflows is essential to protecting an organization's "crown jewels" from sophisticated insider exfiltration.

Summary updated

Abstract

Nowadays, organizations collect vast quantities of sensitive information in `Enterprise Resource Planning' (ERP) systems, such as accounting relevant transactions, customer master data, or strategic sales price information. The leakage of such information poses a severe threat for companies as the number of incidents and the reputational damage to those experiencing them continue to increase. At the same time, discoveries in deep learning research revealed that machine learning models could be maliciously misused to create new attack vectors. Understanding the nature of such attacks becomes increasingly important for the (internal) audit and fraud examination practice. The creation of such an awareness holds in particular for the fraudulent data leakage using deep learning-based steganographic techniques that might remain undetected by state-of-the-art `Computer Assisted Audit Techniques' (CAATs). In this work, we introduce a real-world `threat model' designed to leak sensitive accounting data. In addition, we show that a deep steganographic process, constituted by three neural networks, can be trained to hide such data in unobtrusive `day-to-day' images. Finally, we provide qualitative and quantitative evaluations on two publicly available real-world payment datasets.
BibTeX
@misc{schreyer2020leaking,
  author = {Marco Schreyer and Christian Schulze and Damian Borth},
  title = {{Leaking Sensitive Financial Accounting Data in Plain Sight using Deep Autoencoder Neural Networks}},
  year = {2020},
  eprint = {2012.07110},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2012.07110},
}
Figure