Mitigating Data Injection Attacks on Federated Learning

Or Ohev Shalom, Amir Leshem, Waheed U. Bajwa

Introduction

Federated Learning (FL) has emerged as a powerful paradigm for privacy-preserving machine learning, allowing multiple agents to collaboratively train a shared global model without exchanging raw data. However, this decentralized structure introduces a significant security vulnerability: "data injection" or "poisoning" attacks. In these scenarios, malicious participants can send corrupted local updates to a central coordinating node, potentially sabotaging the global model's performance or inducing specific misclassifications.

Federated Learning architecture with a coordinating node, honest agents, and a potential attacker.

The primary challenge for the coordinating node is its inability to inspect private local datasets. Consequently, it cannot easily distinguish between legitimate updates and malicious ones designed to steer the model toward failure. This paper proposes a robust detection and mitigation scheme based on coordinate-wise medians and temporal majority voting to secure FL against such adversarial behavior.

Methodology: Median-Based Detection and Mitigation

The proposed defense operates at the coordinating node and consists of two primary mechanisms:

  1. Coordinate-Wise Median Metric: Traditional FL aggregation often uses simple averaging, which is highly sensitive to outliers. The authors replace this with a coordinate-wise median. By calculating the median for each model parameter, the coordinator establishes a robust reference point. A "suspicion score" is then assigned to each agent based on the $L_\infty$ norm—the maximum absolute difference—between their update and the calculated median.
  2. Temporal Majority Voting: To prevent the accidental exclusion of honest agents with noisy data, the system employs a temporal window ($\Delta T$). An agent is only excluded from the global model if they are flagged as suspicious in more than 50% of the updates within that window.

A key feature of this approach is its "forgiving" nature. The coordinator continues to evaluate updates from excluded agents; if their contributions realign with the global median, they are reintegrated into the training process. This ensures resilience against transient noise or data heterogeneity while maintaining a firm stance against persistent attackers.

Experimental Evaluation

The researchers benchmarked their method using the MNIST handwritten digit dataset, distributing 60,000 training images across networks of 5, 10, and 20 agents. A 7-layer Convolutional Neural Network (CNN) served as the model architecture. The study focused on two primary attack vectors: Constant-Output (forcing a single prediction) and Label-Flipping (systematic misclassification).

Key Results

The defense demonstrated remarkable efficacy in neutralizing adversarial updates. Without mitigation, a single attacker in a 20-agent network could eventually drive the global classification error to nearly 90%. In contrast, the proposed median-based scheme maintained low error rates across various network sizes.

Classification error remains low and stable when the median-based mitigation scheme is active, even in the presence of an attacker.

The simulations revealed that the coordinator identifies and isolates attackers early in the training cycle. This rapid detection allows the global model to converge to a truthful state as if the malicious actor were never present. Furthermore, the authors provide a mathematical proof using the Borel-Cantelli lemma, demonstrating that the system will eventually ignore all attackers and converge to the correct model with a probability of one, provided the majority of agents remain honest.

Discussion and Future Directions

While the median-based approach is provably secure and computationally efficient for standard benchmarks, several areas for future research remain:

  • Non-i.i.d. Data: The current theoretical framework assumes data is independent and identically distributed (i.i.d.). Real-world FL often involves significant data heterogeneity, which might increase the rate of false positives in the detection scheme.
  • Majority Rule Constraints: The defense relies on the assumption that more than 50% of participants are honest. It remains vulnerable to "51% attacks" where a malicious entity controls the majority of the network.
  • Scalability: Calculating coordinate-wise medians for massive models, such as Large Language Models (LLMs) with billions of parameters, may introduce significant overhead. Future iterations could explore sampling techniques or "sketched" medians to enhance scalability.

In conclusion, this median-based defense offers a robust, mathematically grounded solution to the problem of data injection in Federated Learning, significantly increasing the cost and difficulty for adversaries attempting to compromise shared models.

Summary updated

BibTeX
@inproceedings{shalom2024mitigating,
  author = {Or Ohev Shalom and Amir Leshem and Waheed U. Bajwa},
  title = {{Mitigating Data Injection Attacks on Federated Learning}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
  doi = {10.1109/icassp48485.2024.10446615},
}
Figure