Abstract
Attacks on Federated Learning (FL) can severely reduce the quality of the generated models and limit the usefulness of this emerging learning paradigm that enables on-premise decentralized learning. However, existing untargeted attacks are not practical for many scenarios as they assume that i) the attacker knows every update of benign clients, or ii) the attacker has a large dataset to locally train updates imitating benign parties. In this paper, we propose a data-free untargeted attack (DFA) that synthesizes malicious data to craft adversarial models without eavesdropping on the transmission of benign clients at all or requiring a large quantity of task-specific training data. We design two variants of DFA, namely DFA-R and DFA-G, which differ in how they trade off stealthiness and effectiveness. Specifically, DFA-R iteratively optimizes a malicious data layer to minimize the prediction confidence of all outputs of the global model, whereas DFA-G interactively trains a malicious data generator network by steering the output of the global model toward a particular class. Experimental results on Fashion-MNIST, Cifar-10, and SVHN show that DFA, despite requiring fewer assumptions than existing attacks, achieves similar or even higher attack success rate than state-of-the-art untargeted attacks against various state-of-the-art defense mechanisms. Concretely, they can evade all considered defense mechanisms in at least 50% of the cases for CIFAR-10 and often reduce the accuracy by more than a factor of 2. Consequently, we design REFD, a defense specifically crafted to protect against data-free attacks. REFD leverages a reference dataset to detect updates that are biased or have a low confidence. It greatly improves upon existing defenses by filtering out the malicious updates and achieves high global model accuracy.
Introduction
Federated Learning (FL) allows multiple clients to collaboratively train a machine learning model without sharing their raw, private data. However, this decentralized structure makes FL highly vulnerable to untargeted poisoning attacks, where malicious clients submit corrupted model updates to degrade the global model's overall accuracy.
Existing state-of-the-art untargeted attacks (like LIE, Fang, or Min-Max) rely on strong, often unrealistic assumptions: the attacker must either be able to eavesdrop on the updates sent by benign clients, or they must possess a large, task-specific dataset to mimic normal training behavior. In real-world cross-device FL—such as training a medical model for a rare disease or participating anonymously via Tor—attackers likely have neither. This research investigates whether an attacker can successfully execute an untargeted poisoning attack without any real training data or knowledge of other clients' updates.
Methodology: Data-Free Attacks (DFA)
The proposed Data-Free Attack (DFA) framework synthesizes malicious data using only the globally broadcasted model as a guide. This "zero-knowledge" approach removes the need for local datasets or eavesdropping.
DFA Variants
- DFA-R (Filter-based): This variant aims to destroy the model's confidence. It applies a trainable convolutional filter to random dummy images. The filter is optimized so that the resulting synthetic images force the global model to output equal probabilities across all classes (e.g., 10% confidence for all 10 classes).
- DFA-G (Generator-based): This variant trains a Generative Adversarial Network (GAN)-style generator to synthesize images that the global model confidently misclassifies. It actively steers the model's output away from a randomly chosen target class.
Both variants utilize Distance-based Regularization to evade anomaly-detection defenses. By using a custom loss function, the malicious model updates are forced to remain mathematically close to the previous rounds' global models, ensuring the poisoned updates do not appear as statistical outliers to the server.
Defense via REFD
To counter these attacks, the authors introduce REFD, a novel server-side defense. Unlike traditional defenses (like mKrum or Bulyan) that look at the statistical distance between model weights, REFD uses a small, balanced reference dataset to evaluate how the updated models behave. It calculates a D-score based on two metrics: Balance (is the update biased toward one class?) and Confidence (does the update destroy prediction confidence?). Updates with low D-scores are rejected.
Experimental Results
The experiments were conducted on three standard image classification benchmarks: Fashion-MNIST, CIFAR-10, and SVHN. To simulate real-world FL environments, the training data was distributed among clients heterogeneously (non-IID) using a Dirichlet distribution.
Key Findings
- Data is Optional for Poisoning: DFA achieves similar or higher Attack Success Rates (ASR) compared to SOTA attacks that use real data. On CIFAR-10, DFA-G bypassed the aggressive Bulyan defense 56.6% of the time, reducing the global model's accuracy from 50% to ~21%.
- Stealth via Regularization: By adding a distance-based regularization term, attackers can ensure their malicious updates mimic the statistical trajectory of normal training progress, slipping past robust aggregators.
- Heterogeneity as a Shield: High variance among normal updates in non-IID settings creates a noisy environment where malicious updates easily blend in, making data-free attacks more effective.
- Behavioral Testing Superiority: REFD successfully neutralized DFA attacks where statistical defenses failed. On Fashion-MNIST with highly heterogeneous data, REFD maintained a global model accuracy of >70% under attack, whereas the Bulyan defense collapsed to ~40%.
Conclusion
DFA demonstrates that attackers do not need real data to break Federated Learning. By using the global model as an oracle, they can reverse-engineer synthetic data optimized to degrade model accuracy while remaining stealthy. While traditional statistical defenses struggle against these sophisticated updates, behavioral evaluation via a reference dataset (REFD) provides a robust path forward for securing decentralized learning systems.
Summary updated