MARS: A Malignity-Aware Backdoor Defense in Federated Learning

Wei Wan, Yuxuan Ning, Zhicong Huang, Cheng Hong, Shengshan Hu, Ziqi Zhou, Yechao Zhang, Tianqing Zhu, Wanlei Zhou, Leo Yu Zhang

TL;DR

MARS (Malignity-Aware backdooR defenSe) is a novel defense framework designed to protect Federated Learning systems from state-of-the-art, adaptive backdoor attacks. Instead of relying on easily spoofed statistical metrics like update norms or cosine similarity, MARS calculates a "backdoor energy" metric for individual neurons to uncover malicious intent. By pairing this metric with an order-insensitive Wasserstein distance-based clustering algorithm, MARS accurately filters out poisoned local models without sacrificing global model accuracy.

What Problem Does This Paper Address?

Federated Learning (FL) allows multiple clients to collaboratively train a machine learning model by sharing local model updates (parameters) rather than raw data. While privacy-preserving, this decentralized nature makes FL highly vulnerable to backdoor attacks, where an attacker injects malicious updates to make the global model misclassify specific inputs (e.g., treating a stop sign with a sticker as a speed limit sign).

Existing defenses typically try to catch these attacks using empirical statistical measures—such as checking if an update's norm is too large, or if its distribution looks like an outlier. However, the authors demonstrate that modern, adaptive backdoor attacks (like 3DFed, CerP, and DarkFed) can easily bypass these defenses. Attackers can simply add constraints or noise masks during local training so that their malicious updates perfectly mimic the statistical properties, norms, and distributions of benign updates.

The core issue is that existing defenses are loosely coupled with the actual backdoor behavior; they look at how the model parameters look on the surface rather than evaluating the model's internal structural integrity.

Methods

MARS introduces a defense strategy deployed at the central server that operates in three main phases:

  1. Calculating Backdoor Energy (BE): Backdoor attacks inherently establish "shortcuts" within a neural network, which abnormally alters the activation behavior of certain neurons. MARS quantifies this abnormality as "backdoor energy" using the Lipschitz constant of each neuron. Crucially, because the server cannot access clean data or the attacker's trigger, MARS approximates this data-freely using the spectral norm of the model's weight matrices.
  2. Forming Concentrated Backdoor Energy (CBE): Since backdoors only manipulate a small subset of neurons, MARS extracts the top $\kappa\%$ (typically 5%) highest BE values from each layer and concatenates them into a single vector. This isolates the backdoor information and filters out benign noise.
  3. Wasserstein Distance-Based Clustering (K-WMeans): The server must cluster these CBE vectors to separate benign and malicious models. Standard clustering techniques use Euclidean or Cosine distance, which are highly sensitive to element order. In FL, different clients might learn the backdoor on entirely different neurons, meaning the high-energy values will appear in different positions in the vector. To solve this, MARS uses Wasserstein distance, which evaluates the overall probability density rather than the exact coordinate order.
Figure 1: Overview of MARS. To facilitate understanding, we simplify the FL system to include only four clients. The first and third clients upload backdoor models, while the second and fourth clients upload benign models. Red circles represent higher backdoor energy, while green circles indicate lower backdoor energy.

Finally, MARS selects the cluster with the smaller center norm to aggregate into the global model, as malicious models inherently exhibit higher backdoor energy. If the distance between the two clusters is below a threshold $\epsilon$, it assumes no attack is occurring and aggregates both clusters to preserve model convergence.

Datasets

The authors evaluated MARS across several standard computer vision and natural language processing benchmarks under non-IID (heterogeneous) data distributions:

  • MNIST (with a standard CNN)
  • CIFAR-10 and CIFAR-100 (using ResNet-18)
  • ImageNet (large-scale evaluation using ReXNet)
  • IMDB (NLP task using an LSTM)

Key Insights and Ideas

  • The "Order-Sensitivity" Trap: Traditional distance metrics fail in distributed clustering because they expect features to align perfectly by index. The authors show that Wasserstein distance acts as an order-insensitive metric, making it uniquely suited for identifying malicious updates across independent clients who have altered different internal neurons.
Metric (L1, L2) Backdoor/Backdoor (L1, L3) Backdoor/Benign (L2, L3) Backdoor/Benign
Euclidean 6.16 5.48 6.16
Cosine 0.31 0.10 0.07
Wasserstein 0.40 2.00 2.40

As shown in the paper's toy example, Euclidean and Cosine metrics mistakenly group a backdoor model (L1 or L2) closer to a benign model (L3). Wasserstein distance correctly identifies the two backdoor models as highly similar despite their structural permutations.

  • Malignity-Aware vs. Statistical Defenses: Surface-level parameters (like weight norms) are easy for an attacker to manipulate. Internal mathematical properties (like Lipschitz smoothness bounds) are tightly coupled with the function the network is computing, making it incredibly difficult for an attacker to maintain a high backdoor success rate while hiding their structural footprint.

Key Results

  • Defeating Advanced Attacks: Against the state-of-the-art 3DFed attack on CIFAR-10, standard defenses like FLAME or FoolsGold saw their Comprehensive Ability of Defense (CAD) drop to ~33% and ~46% respectively. MARS maintained a 93.80% CAD, keeping the Attack Success Rate (ASR) under 10%.
  • Robustness to High Attacker Ratios: MARS does not assume that benign clients are the majority. In experiments varying the attacker ratio from 0% up to 95%, MARS consistently achieved a 100% True Positive Rate (TPR) in identifying malicious clients and a 0% False Positive Rate (FPR).
  • Low Computational Overhead: By restricting BE calculations to convolutional and batch-normalization layers and optimizing via dot products, MARS completes its server-side aggregation in 6.57 seconds on CIFAR-10, which is comparable to basic FedAvg (2.07s) and significantly faster than other inspection defenses like DeepSight (101.69s).

Limitations and Future Work

While highly effective against backdoor injections, MARS is specifically tailored to threats that alter internal neuron smoothness. The authors note the following limitations:

  • Alternative Threat Models: MARS is not designed to detect non-backdoor threats, such as "free-rider" attacks (where clients contribute lazy/empty updates without degrading performance) or privacy-stealing attacks (like gradient inversion).
  • Fully Connected Layer Omission: To maintain rapid runtime, MARS omits fully connected (FC) layers from its detection pipeline. Although the authors proved that executing a backdoor attack solely through FC layers yields poor stealth and low effectiveness, sophisticated hybrid partial-layer attacks remain an area for ongoing scrutiny.

Summary updated

Abstract

Federated Learning (FL) is a distributed paradigm aimed at protecting participant data privacy by exchanging model parameters to achieve high-quality model training. However, this distributed nature also makes FL highly vulnerable to backdoor attacks. Notably, the recently proposed state-of-the-art (SOTA) attack, 3DFed (SP2023), uses an indicator mechanism to determine whether the backdoor models have been accepted by the defender and adaptively optimizes backdoor models, rendering existing defenses ineffective. In this paper, we first reveal that the failure of existing defenses lies in the employment of empirical statistical measures that are loosely coupled with backdoor attacks. Motivated by this, we propose a Malignity-Aware backdooR defenSe (MARS) that leverages backdoor energy (BE) to indicate the malicious extent of each neuron. To amplify malignity, we further extract the most prominent BE values from each model to form a concentrated backdoor energy (CBE). Finally, a novel Wasserstein distance-based clustering method is introduced to effectively identify backdoor models. Extensive experiments demonstrate that MARS can defend against SOTA backdoor attacks and significantly outperforms existing defenses.
BibTeX
@inproceedings{wan2025mars,
  author = {Wei Wan and Yuxuan Ning and Zhicong Huang and Cheng Hong and Shengshan Hu and Ziqi Zhou and Yechao Zhang and Tianqing Zhu and Wanlei Zhou and Leo Yu Zhang},
  title = {{MARS: A Malignity-Aware Backdoor Defense in Federated Learning}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=3kmbucBZPA},
}
Figure