Enhancing the Effectiveness and Durability of Backdoor Attacks in Federated Learning through Maximizing Task Distinction

Zhaoxin Wang, Handing Wang, Cong Tian, Yaochu JinarXiv 2025 · OpenReview

Abstract

Federated learning allows multiple participants to collaboratively train a central model without sharing their private data. However, this distributed nature also exposes new attack surfaces. In particular, backdoor attacks allow attackers to implant malicious behaviors into the global model while maintaining high accuracy on benign inputs. Existing attacks usually rely on fixed patterns or adversarial perturbations as triggers, which tightly couple the main and backdoor tasks. This coupling makes them vulnerable to dilution by honest updates and limits their persistence under federated defenses. In this work, we propose an approach to decouple the backdoor task from the main task by dynamically optimizing the backdoor trigger within a min-max framework. The inner layer maximizes the performance gap between poisoned and benign samples, ensuring that the contributions of benign users have minimal impact on the backdoor. The outer process injects the adaptive triggers into the local model. We evaluate our method on both computer vision and natural language tasks, and compare it with six backdoor attack methods under six defense algorithms. Experimental results show that our method achieves good attack performance and can be easily integrated into existing backdoor attack techniques.

Introduction

Federated Learning (FL) has emerged as a powerful paradigm for privacy-preserving machine learning, enabling multiple parties to train a shared model without exchanging raw data. Despite its benefits, the decentralized nature of FL introduces significant security vulnerabilities, most notably backdoor attacks. In these attacks, a malicious participant poisons their local updates to ensure the global model performs a specific adversarial action when triggered by a specific input pattern, while otherwise appearing benign.

A critical challenge for such attacks is durability. In a typical FL environment, updates from numerous honest participants tend to "dilute" the malicious task over time—a phenomenon known as catastrophic forgetting. Most current backdoor attacks are fragile, losing effectiveness shortly after the attacker ceases participation. This paper introduces EDBA (Effectiveness and Durability of Backdoor Attacks), a framework designed to create persistent backdoors by maximizing the mathematical distinction between the main task and the backdoor task.

Methodology: Task Distinction through Min-Max Optimization

The core philosophy of EDBA is Task Distinction. By decoupling the backdoor task from the main task, the updates intended for the benign objective do not interfere with the weights responsible for the malicious behavior.

EDBA implements this via a dynamic min-max optimization framework:

  1. Inner Maximization (Trigger Generation): Rather than using static triggers, EDBA dynamically optimizes the trigger using gradient ascent. It seeks to maximize the distance (e.g., via cosine similarity) between the model's internal representations of clean and poisoned samples. This ensures the model perceives poisoned data as fundamentally unique.
  2. Outer Minimization (Backdoor Injection): With the distinct trigger identified, the attacker trains the local model to achieve high accuracy on both the main and backdoor tasks.

For Computer Vision tasks, the framework optimizes pixel values. For Natural Language Processing (NLP), it utilizes a "position importance ranking" to insert rare trigger words into sentences without altering their original semantic meaning.

Experimental Evaluation

The effectiveness of EDBA was evaluated across several standard benchmarks:

  • Computer Vision: MNIST, CIFAR-10, CIFAR-100, and Tiny-ImageNet.
  • NLP: Yelp-Review and IMDB sentiment analysis.

Tests were conducted in both IID and Non-IID settings, the latter being more representative of real-world FL deployments where data distribution varies significantly across users.

Key Results and Discussion

EDBA demonstrates superior performance in terms of both success rate and durability. On the CIFAR-10 dataset in a Non-IID setting, the attack achieved a 95.71% Attack Success Rate (ASR).

Figure 2: Attack performance across various computer vision datasets and compromised ratios. Subfigure (a) shows rapid convergence on CIFAR10, while (b) demonstrates high ASR on Tiny-ImageNet even with a low compromised ratio.

The attack is also highly efficient, requiring as few as 5% compromised clients to be effective. In NLP tasks, such as the Yelp-Review dataset, the backdoor task reaches near-perfect accuracy almost immediately.

Figure 3: Performance of EDBA on the Pool-Yelp NLP task, showing the rapid and stable injection of the backdoor task alongside the main task.

Perhaps most significantly, EDBA exhibits extreme durability. In lifespan tests on CIFAR-10, while baseline attacks like BadNets were quickly washed out by honest updates, EDBA maintained a high ASR for over 1,000 additional rounds after the attacker stopped contributing. Furthermore, because EDBA-poisoned updates are statistically similar to benign ones, they successfully bypass robust aggregation defenses such as Krum, Median, FLAME, and Freqfed.

Conclusion

EDBA addresses the fundamental fragility of backdoor attacks in Federated Learning by mathematically decoupling the malicious task from the benign one. By dynamically optimizing triggers to maximize task distinction, the attack achieves unprecedented durability and resistance to common defenses. While the computational overhead for trigger optimization is a current limitation, the results underscore a significant security gap in existing FL frameworks that may require new defensive strategies, such as retrospective model rollbacks.

Summary updated

Abstract

Federated learning allows multiple participants to collaboratively train a central model without sharing their private data. However, this distributed nature also exposes new attack surfaces. In particular, backdoor attacks allow attackers to implant malicious behaviors into the global model while maintaining high accuracy on benign inputs. Existing attacks usually rely on fixed patterns or adversarial perturbations as triggers, which tightly couple the main and backdoor tasks. This coupling makes them vulnerable to dilution by honest updates and limits their persistence under federated defenses. In this work, we propose an approach to decouple the backdoor task from the main task by dynamically optimizing the backdoor trigger within a min-max framework. The inner layer maximizes the performance gap between poisoned and benign samples, ensuring that the contributions of benign users have minimal impact on the backdoor. The outer process injects the adaptive triggers into the local model. We evaluate our method on both computer vision and natural language tasks, and compare it with six backdoor attack methods under six defense algorithms. Experimental results show that our method achieves good attack performance and can be easily integrated into existing backdoor attack techniques.
BibTeX
@misc{wang2025enhancing,
  author = {Zhaoxin Wang and Handing Wang and Cong Tian and Yaochu Jin},
  title = {{Enhancing the Effectiveness and Durability of Backdoor Attacks in Federated Learning through Maximizing Task Distinction}},
  year = {2025},
  eprint = {2509.18904},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.18904},
}
Figure