Abstract
Federated Learning (FL) is a collaborative machine learning approach allowing participants to jointly train a model without having to share their private, potentially sensitive local datasets with others. Despite its benefits, FL is vulnerable to backdoor attacks, in which an adversary injects manipulated model updates into the model aggregation process so that the resulting model will provide targeted false predictions for specific adversary-chosen inputs. Proposed defenses against backdoor attacks based on detecting and filtering out malicious model updates consider only very specific and limited attacker models, whereas defenses based on differential privacy-inspired noise injection significantly deteriorate the benign performance of the aggregated model. To address these deficiencies, we introduce FLAME, a defense framework that estimates the sufficient amount of noise to be injected to ensure the elimination of backdoors while maintaining the model performance. To minimize the required amount of noise, FLAME uses a model clustering and weight clipping approach. Our evaluation of FLAME on several datasets demonstrates that FLAME removes backdoors effectively with a negligible impact on the benign performance of the models. Furthermore, we provide a comprehensive analysis of recent attempts to circumvent FLAME (e.g., 3DFed), showing that our defense mechanism effectively repels these attempted attacks.
The Vulnerability of Federated Learning
In Federated Learning, a central server coordinates training across thousands of user devices (clients) without ever seeing their raw data. While this architecture preserves privacy, it creates a significant security gap: "backdoor attacks." An attacker can compromise a subset of clients to inject "poisoned" updates. These updates train the model to behave normally on standard inputs but fail predictably when encountering a specific "trigger"—such as a specific pixel pattern in an image or a specific phrase in a text.
Existing defenses typically fall into two categories: anomaly detection and differential privacy (DP). Anomaly detection often fails against "stealthy" attacks that mimic benign behavior, while DP-based noise injection often adds so much noise that the model's utility for its primary task is destroyed. FLAME bridges this gap by providing a robust defense that preserves model utility.
The FLAME Defense Pipeline
FLAME (Federated Learning Adaptive Model Ensemble) introduces a three-stage pipeline designed to "tame" backdoors by combining filtering, magnitude control, and noise injection.
1. Dynamic Model Filtering (Clustering)
Instead of assuming a fixed number of attackers, FLAME utilizes HDBSCAN, a density-based clustering algorithm. It calculates the cosine distance (the angle) between model updates. Because benign updates are generally moving toward a similar optimization goal, they tend to cluster together. Malicious updates that deviate significantly are flagged as outliers and removed from the aggregation process.
2. Adaptive Weight Clipping
Attackers often "scale up" the magnitude of their updates to maximize their impact on the global model. FLAME thwarts this by clipping the $L_2$-norm (magnitude) of all updates. Crucially, the clipping bound is adaptive—it is set to the median magnitude of all updates in a given round. This ensures that no single client can exert undue influence, and the defense remains relevant as model updates naturally shrink during convergence.
3. Adaptive Noising
This final "cleanup" step adds a calculated amount of Gaussian noise to the aggregated model. The synergy between the stages is vital: because the updates were already clustered and clipped, the "sensitivity" of the model is significantly lower. This allows FLAME to use a mathematically proven minimal amount of noise to eliminate any surviving backdoor influence without destroying the main task's accuracy.
Empirical Performance and Robustness
The authors evaluated FLAME across three distinct domains: Word Prediction (Reddit dataset), Image Classification (CIFAR-10, MNIST, Tiny-ImageNet), and IoT Intrusion Detection (IoT-Traffic).
Key Results: * Backdoor Elimination: In nearly all scenarios, FLAME reduced the Backdoor Accuracy (BA)—the success rate of the attack—to near 0%. * Preserving Utility: FLAME had a negligible impact on Main Task Accuracy (MA). For instance, on CIFAR-10, the benign model's 92.2% accuracy only dropped to 91.9% under attack with FLAME active. * Resilience to Adaptive Attacks: A significant portion of the research addresses recent papers (like 3DFed) claiming to bypass FLAME. The authors demonstrate that these attacks only appear successful when they ignore the multi-layered nature of the defense; while an attacker might attempt to bypass clustering, they are ultimately neutralized by the subsequent clipping and noising steps.
Conclusion
FLAME represents a significant advancement in secure Federated Learning. By treating clustering and clipping as preprocessing steps for noise injection, it achieves a balance between security and utility that previous methods could not reach. While Private FLAME (using Secure Multi-Party Computation) introduces computational overhead, the core framework provides a robust, adaptive shield against the evolving landscape of backdoor threats.
Summary updated