# SAB:A Stealing and Robust Backdoor Attack based on Steganographic Algorithm against Federated Learning

**Authors:** Weida Xu, Yang Xu, Sicong Zhang  
**arXiv:** [2408.13773](https://arxiv.org/abs/2408.13773)  
**Date:** 2024-08-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.13773v1/

## Key points

- Federated learning is an innovative network architecture in technology aimed at safeguarding user privacy.
- The increasing adoption of federated learning has exposed vulnerabilities to backdoor attacks within the system.
- There is a growing need to explore and enhance security measures in federated learning.
- Current research has identified various backdoors that can be detected through human inspection or countered by detection algorithms.
- A new approach called SAB (Stealing and Robust Backdoor Attack based on Steganographic Algorithm against Federated Learning) has been introduced to address backdoor attacks in federated learning.
- SAB uses steganographic algorithms, particularly image steganography, to construct triggers that enhance the accuracy of backdoors.
- SAB incorporates multiple loss joint computation techniques for generating triggers with smaller distances to benign samples and greater imperceptibility to the human eye.
- SAB can mitigate or evade specific backdoor defense methods, making it a powerful tool for malicious actors exploiting vulnerabilities in federated learning systems.
- The bottom-95% method used in SAB extends the lifespan of backdoor attacks by updating gradients on minor value points, reducing detection likelihood.
- SAB integrates Sparse-update techniques to improve generalization of backdoors and overall accuracy in federated learning environments.

## Summary

In the realm of technology, federated learning has emerged as an innovative network architecture aimed at safeguarding user privacy. However, the increasing adoption of federated learning has also brought to light the vulnerability of backdoor attacks within this system. As a result, there is a growing need to explore and enhance the security measures in federated learning. Current research on federated learning has identified various backdoors that can be easily detected through human inspection or countered by detection algorithms. In response to this challenge, a new objective has been established to develop more sophisticated and resilient backdoor attacks in federated learning. This led to the introduction of a novel approach known as SAB (Stealing and Robust Backdoor Attack based on Steganographic Algorithm against Federated Learning). SAB is specifically tailored for addressing backdoor attacks in federated learning by introducing an alternative gradient updating mechanism. The attack leverages steganographic algorithms, particularly image steganography, to construct full-size triggers that enhance the accuracy of backdoors. Additionally, SAB utilizes multiple loss joint computation techniques to generate triggers with smaller distances to benign samples and greater imperceptibility to the human eye. One notable feature of SAB is its ability to mitigate or evade specific backdoor defense methods, making it a formidable tool for malicious actors seeking to exploit vulnerabilities in federated learning systems. The bottom-95% method employed in SAB extends the lifespan of backdoor attacks by updating gradients on minor value points, reducing the likelihood of being detected and cleaned. Furthermore, SAB incorporates Sparse-update techniques to enhance the generalization of backdoors and improve overall accuracy. By combining advanced steganographic algorithms with innovative gradient updating mechanisms, SAB represents a significant advancement in combating backdoor attacks in federated learning environments. Authors Weida Xu, Yang Xu, and Sicong Zhang have contributed significantly to this research effort, highlighting the importance of developing robust security measures in federated learning systems. Their work underscores the ongoing evolution and refinement of strategies aimed at protecting user privacy and data integrity in modern technological landscapes.